I architect production AI systems
for enterprise.
NVIDIA Certified AI Architect. Founder of BrainyNeurals. Since 2018, I have led the delivery of 70+ production AI systems across manufacturing, banking and financial services, healthcare, logistics, and construction. My work focuses on the hard part: making AI systems actually ship to production and continue performing there.
-
01 · Certification
NVIDIA Certified
AI Architect -
02 · Tenure
8+ Years
Applied AI -
03 · Output
70+ Production
Systems Delivered -
04 · Role
Founder ·
BrainyNeurals
A brief introduction — in my own words.
I started working with AI in 2018, during what I now recognize as the last year before the discipline became fashionable. At the time, I was a year into my career as an embedded systems engineer. I wrote C and C++ firmware. I had a B.Tech in Electronics and Communication and an M.Tech in Embedded Systems. I was supposed to be on a safe and predictable path.
I walked away from it because I spent six weeks teaching myself NVIDIA’s DeepStream SDK and YoloV2, and by the end of those six weeks I could see what the next decade would look like.
Eight years later, I have led the delivery of more than 70 production AI systems. I have made a lot of mistakes in that time. Most of them were architectural, because that is where AI mistakes compound most expensively. I have learned to design systems that actually ship — not systems that demo well.
This page exists because people Google me before they hire BrainyNeurals, and I would rather give them the substantive version than let LinkedIn tell the abbreviated one. If you scrolled here from a podcast, a conference bio, or a LinkedIn message, the rest of the page is the long form.
If you are evaluating BrainyNeurals for an enterprise AI engagement, I am the person you will work with — or at minimum, the person who will personally review the engagement. My email is at the bottom of this page. I read it myself.
— Mitesh
Why I bet my career on AI in 2018 — and what that bet has taught me.
The context.
In 2018, the word “AI” was already in the industry vocabulary, but applied AI was still a niche concern. Deep learning was taught in universities; it was not yet running in factories. Convolutional networks existed; they were not yet inspecting your tires. The tooling was primitive by 2026 standards — we had TensorFlow 1.x, early PyTorch, NVIDIA DeepStream had just been released, and most “AI” deployments in industry were actually classical computer vision with a thin CNN layer bolted on.
I was a year into firmware engineering. I had written bootloaders. I had debugged interrupt handlers on microcontrollers. I had the safe, predictable career of an embedded engineer. My B.Tech was in Electronics and Communication; my M.Tech was in Embedded Systems. The reasonable next step was a senior embedded role at a larger firm.
The six weeks that changed the trajectory.
I came across NVIDIA’s DeepStream SDK while looking for a way to run real-time video analytics on embedded hardware. DeepStream was — and still is — a framework for processing video streams on NVIDIA GPUs. I downloaded it, ran the tutorials, and within a week I had a YoloV2 object detector running on real video feeds.
I spent the next six weeks doing this after work. Every evening, I would come home from firmware engineering and work on computer vision. I learned CUDA basics. I learned why inference latency matters. I learned how to convert models between frameworks. I learned that the hard part of production computer vision is not training the model — it is everything downstream of the model.
The thesis I formed.
By the end of those six weeks, I had formed a thesis that I am now confident enough to state as a prediction proven correct: AI would not be a feature that software firms bolt onto existing services. AI would be a separate discipline, with its own architectural patterns, its own failure modes, its own operational disciplines. The engineers who committed to it in 2018 would compound eight years of specialist expertise by the time enterprises needed production systems in 2026. The engineers who treated it as a side service would still be learning.
I decided to commit. I resigned from my firmware role. I spent most of 2018 doing independent AI project work — sports analytics, construction safety, retail counting — proving to myself that the thesis held. It did.
Why I founded BrainyNeurals.
I founded BrainyNeurals because I could see the inverse of my thesis forming in the market: most software consultancies would eventually try to add AI to their service menu, and most would fail at it. Not because they lacked talent, but because they would lack the depth of experience that only comes from doing nothing else for eight years. And enterprises building their first or second AI system cannot afford a partner who is learning on their engagement. The cost of mid-engagement amateurism — in money, in internal credibility for the AI program, in lost time — is always higher than the cost of engaging a specialist from the start.
That inverse thesis is now visible in the market everywhere. The generalist consultancies that added AI practices three years ago are still learning. The specialist AI-only firms that started in 2018 compounded eight years of experience. The gap widens rather than closes. BrainyNeurals was built to occupy the specialist side of that gap.
What the bet taught me.
Two things stand out. First: architectural decisions made early in an AI engagement compound disproportionately. A bad model trained on good architecture can be retrained; a good model running on bad architecture cannot be rescued. This is why every engagement I lead begins with architecture, not code.
Second: specialism is not just a marketing claim. It is a structural hiring, staffing, and scope-control discipline. Maintaining it requires saying no to adjacent service lines even when clients want them. We have said no to web development, mobile app development, and generic software consulting requests many times. Those are real revenue opportunities we have turned down. Each turn-down has strengthened the specialism. If we had said yes, we would now be a 30-person generalist with a weak AI practice instead of a 20-person specialist with deep AI.
- 2014B.Tech, Electronics & Communication Engineering
- 2016M.Tech, Embedded Systems
- 2017Production firmware engineering role (C/C++)
- 2018Six-week DeepStream deep-dive → AI pivot decision → BrainyNeurals founded
- 2020First enterprise CV engagement delivered
- 2022ISO 27001 certification achieved
- 202470+ production AI systems milestone reached
- 2026NVIDIA Certified AI Architect credential attained
My role at BrainyNeurals is not ceremonial.
Most founder-CEOs at consultancies of our size are split between technical work and commercial work. I am no exception. Pretending otherwise would undersell the reality — I do both because specialist firms of 20 people cannot afford two leaders, and also because the commercial side of this firm is where I can most directly reinforce the specialist positioning that the technical side executes.
A clean honest picture of what I actually do each week
The Commercial Side
I own the first-contact conversation for most enterprise inquiries. When the contact form submits or an email arrives at mitesh@brainyneurals.com, I am usually the first person to read it. I decide whether it is a fit, I write the initial response, and I run the first call for engagements above a certain threshold.
Specifically:
- Client discovery calls — I personally run the 30-minute initial call for most inbound inquiries
- Proposals and scoping documents — I write or review every engagement proposal before it goes out
- Sales strategy — inbound, outbound, account-based outreach patterns, partner network development
- AI thought leadership publishing — technical blog posts, LinkedIn newsletter, conference talks
- Marketing strategy for AI services — positioning, content strategy, campaign definition
- Client relationship management for strategic accounts
I do this side of the work because in a 20-person firm, the founder is the most credible commercial voice. When an enterprise CTO wants to understand how we would approach their engagement, they want that conversation with the person who has architected 70+ systems — not with a sales development representative.
The Technical Side
On the technical side, I personally review architecture decisions on every engagement above scope threshold. I do not write code for most active engagements — the team does — but I am involved in architectural choices, technical escalations, and post-deployment reviews.
Specifically:
- Architecture review — I sign off on the architecture document produced in every engagement’s scoping phase
- Technical escalations — when the team encounters a design problem, I am the technical escalation point
- Engagement-level production incidents — for any production issue on a system we shipped, I am involved in the post-mortem
- Tool and framework decisions — I review and approve all new framework adoption (new inference servers, new model classes, new deployment targets)
- R&D investment decisions — what the team spends non-billable time building (agentic systems patterns, multimodal architectures, edge fleet tooling)
- Hiring and technical interview panels — I personally interview every senior engineer hire
What I Do Less Of
I no longer write most of the production code. This is not because I cannot — I still code actively for R&D and architectural prototyping. It is because production delivery is better when senior engineers own their code end-to-end without founder interference. I would rather review the architecture than write the implementation.
I also do not run the day-to-day project management of engagements. Senior engineers own engagement delivery; I stay involved strategically without being a project manager.
If you work with us, you will see me in the commercial and architectural moments of our engagement — and you will see senior engineers in the delivery moments. That split is deliberate. It is how a 20-person firm gets the benefit of founder-level seniority without founder-bottleneck delivery.
The technical stack I have actually shipped production systems on.
This section lists tools, frameworks, and platforms I have used in production work — not every technology I have heard of. The distinction matters for enterprise buyers doing due diligence.
Primary competency area — my deepest experience and the area where BrainyNeurals has delivered the most engagements. Production work across object detection, segmentation, classification, depth sensing, multi-camera tracking, and OCR.
Deep specialization — one of the areas where my firmware background converts directly to AI expertise. Production deployments across multiple edge chip families, each with distinct optimization characteristics.
Production model serving infrastructure — the less-visible but more-critical half of AI delivery. This is where most AI engagements fail in practice.
More recent area of expertise (2022+) — production RAG systems and fine-tuned LLM applications. Active R&D investment area for the firm.
Where my firmware background compounds — systems that combine AI with real sensor hardware. Production work across automotive-adjacent, robotics, and industrial deployments.
Non-tool expertise — the consulting-side skills that matter as much as the technical ones for founder-led engagements.
- AI readiness assessment — diagnosing whether an organization can actually absorb a production AI system
- Use case identification and prioritization — which AI use cases actually produce business value vs. which look good in slides
- Architecture review for in-flight engagements led by other teams
- Vendor selection advisory (when we recommend other firms for use cases outside our specialism)
- Build-vs-buy analysis for enterprise AI capability development
Industries where I have personally shipped production AI.
Every industry below is one where I have led engagement delivery — not just where the firm has capability. The distinction matters for founder-involvement claims.
-
01
Manufacturing & Industrial
Tire manufacturing CV inspection — 99.2% defect detection accuracy at 200+ tires/hour on NVIDIA Jetson
Quality inspection, dimensional QC via depth sensing, factory-floor safety monitoring, predictive maintenance
-
02
Banking, Financial Services & Insurance
Document AI system processing 50,000+ documents per month across 47 formats
Document intelligence (IDP), KYC automation, insurance claims processing, fraud pattern detection, RAG systems for internal knowledge bases
-
03
Healthcare
Medical coding automation reducing 48-hour turnaround to 4 hours, HIPAA-compliant with Epic integration
Medical imaging AI, clinical documentation AI, medical coding, healthcare RAG systems for clinical knowledge
-
04
Logistics & Supply Chain
Warehouse safety monitoring CV across distributed facilities
Warehouse CV (safety, inventory, staging), fleet routing optimization, demand forecasting, last-mile delivery AI
-
05
Construction & Infrastructure
Civil engineering plan approval AI reducing 3-week cycle to 4 days
Construction document AI, site safety monitoring via CV, plan review automation, bid document processing
-
06
Sports Analytics
Real-time multi-camera player tracking and heatmap generation for a global sports league
Real-time video analytics, multi-camera tracking, player and ball detection, tactical analysis AI
If your industry is not on this list, it does not automatically mean we are the wrong fit — BrainyNeurals has capability in additional industries. But if I have not personally shipped in your industry, I will say so in our first conversation, and senior engineers with deeper industry-specific context will lead the engagement.
The verifiable record.
Professional Certifications
NVIDIA Certified AI Architect
- Issuing body
- NVIDIA Corporation
- What it is
- A technical certification validating advanced expertise in architecting AI systems on NVIDIA infrastructure. Requires passing a rigorous examination covering AI system design, GPU optimization, model deployment, and production operationalization.
- Significance
- Typically the most rigorous individual AI certification referenced by enterprise procurement teams. Held by a small population of AI architects globally.
- Verification
- Credential ID available on request; LinkedIn profile shows verified badge.
Education
M.Tech in Embedded Systems
- Year
- 2016
- Institution
- To confirm
- Relevance
- The embedded systems graduate training directly supports edge AI specialization — memory-constrained inference, RTOS integration, firmware-to-AI bridges.
B.Tech in Electronics & Communication Engineering
- Year
- 2014
- Institution
- To confirm
- Relevance
- Signal processing, embedded programming, and hardware-software interface training — foundational for the sensor-fusion and edge-deployment work I now do.
Industry Recognition
Upwork Top Rated Plus
- Platform
- Upwork
- Designation
- Top Rated Plus (both individual profile and BrainyNeurals agency profile).
- Significance
- Held continuously since the Top Rated Plus tier was introduced. Represents the top tier of talent on the platform based on client satisfaction and engagement quality metrics.
NVIDIA Inception Partner
- Program
- NVIDIA Inception (via BrainyNeurals).
- Significance
- Active participant in NVIDIA’s global AI accelerator program. Grants access to NVIDIA technical resources, pre-release hardware, and co-marketing opportunities.
AWS Activate + Microsoft for Startups
- Programs
- AWS Activate, Microsoft for Startups (via BrainyNeurals).
- Significance
- Active participation in both major cloud providers’ technology company programs. Enables direct architecture support from cloud solutions architects on client engagements.
Any of the above credentials can be independently verified on request during an enterprise engagement. For NVIDIA Architect, credential ID is provided; for Upwork, profile URL with verified badge; for educational credentials, transcripts available under NDA for procurement due diligence.
Seven engagements I have personally led — at a technical depth that matters.
Seven signature engagements below. Each is anonymized at client level where NDA requires, but detailed at architectural and outcome level. These are engagements I personally led or architected — not firm-wide portfolio claims. More detail available on request under NDA. See the engagements I’ve led →
Tire Manufacturing CV Inspection
Real-time defect detection at 200+ tires per hour on edge hardware.
The challenge was straightforward to state and difficult to execute: inspect every tire coming off the production line for surface defects, at production line speed, using computer vision. The client had previously attempted this with a generalist consultancy; that attempt failed because the inference latency could not keep up with line speed, and deploying a GPU server per line was cost-prohibitive.
The architectural approach I led: NVIDIA Jetson AGX deployment at each inspection station, YOLO-family object detection trained on a curated defect dataset, TensorRT optimization for INT8 inference, custom image preprocessing pipeline to handle the conveyor motion blur. The critical architectural decision was to run inference at the edge rather than centralize — which required the model to be small enough to fit in Jetson memory while still achieving the required accuracy.
99.2% defect detection accuracy in production, 200+ tires per hour throughput maintained, ROI within 8 months based on reduction in defective product reaching customers.
Enterprise Document AI for Financial Services
50,000+ documents per month across 47 formats — 80% reduction in manual review.
The challenge: the client processed tens of thousands of documents per month (contracts, KYC forms, insurance claims, internal memos) spanning 47 distinct document formats. Manual review was a six-person full-time team. An off-the-shelf IDP solution had been evaluated and rejected because it could not handle the format diversity.
The architectural approach: layered document AI pipeline combining layout detection, table extraction, OCR for scanned documents, and LLM-based semantic extraction. Each of the 47 document types got its own post-processing adapter, but all flowed through the same base pipeline. RAG-based knowledge layer for handling edge cases.
80% reduction in manual review time, processing throughput scaled to 50,000+ documents per month, human reviewers reallocated to edge cases and quality control instead of baseline extraction.
Where I publish — and what I am currently thinking about.
This section is updated continuously as I publish. If you are looking for my current thinking on a specific topic and cannot find it below, reach out directly — most of my written commentary happens on LinkedIn and the BrainyNeurals newsletter.
Post title will populate from the CMS feed in Phase 2.
Excerpt will populate from the CMS feed in Phase 2.
Post title will populate from the CMS feed in Phase 2.
Excerpt will populate from the CMS feed in Phase 2.
Post title will populate from the CMS feed in Phase 2.
Excerpt will populate from the CMS feed in Phase 2.
Post title will populate from the CMS feed in Phase 2.
Excerpt will populate from the CMS feed in Phase 2.
Speaking & Podcast Appearances
LinkedIn Featured Posts
Active Research Areas
Agentic AI for Production Workflows
The gap between demo-quality agentic systems and production-grade agentic systems is significant. I am actively building production patterns for agents with durable state, human-in-loop escalation, and audit trails. Published thinking coming Q2–Q3 2026.
Edge Fleet Orchestration at Scale
Edge AI has been my technical core since 2018. The 2026–2028 shift is from single-site deployment to fleet-scale. I am investing R&D time in centralized model distribution, per-site drift monitoring, and bandwidth-efficient update patterns. Published thinking expected throughout 2026.
Multimodal Production Architecture
Production systems fusing vision and language at inference layer rather than application layer. This is where new enterprise use cases will unlock. Early client engagements serve as validation ground. Published thinking expected late 2026 through 2027.
Five positions I hold about production AI — with reasoning, not slogans.
These are my personal positions, not marketing claims. I hold each of them strongly enough that they shape hiring, staffing, engagement decisions, and sometimes client rejection. If you find any of them disagreeable, we may not be the right match — and that is useful information for both of us.
Architecture Before Code, Always.
The most expensive mistakes in AI systems are architectural, not implementation-level. A bad model trained on good architecture can be retrained in a week. A good model running on bad architecture cannot be rescued without a rebuild. This is not theoretical — I have watched other firms’ engagements fail for exactly this reason, and I have had to tell clients that their prior vendor’s work needs to be discarded rather than extended.
Every engagement I lead begins with an architecture phase. Sometimes clients push back because they want to “just start building.” I tell them the truth: the two weeks spent on architecture compresses six months of potential rework downstream. That trade-off is obvious once you have made the inverse mistake twice. I have made it twice. I do not need to make it again.
The Specification is Half the Engagement.
The specifications of an AI system are where most engagements go wrong. Clients often arrive with a specification written by someone who does not understand where the hard parts will be — which data is actually available, which edge cases matter, which accuracy thresholds are business-critical versus nice-to-have. If the specification is wrong, even a perfectly-executed engagement fails.
This is why I personally run the first scoping call on every engagement. I cannot outsource that conversation. Sending a sales development representative to write specifications is where the rot starts — the resulting specification always misses the hard parts, because the SDR cannot identify them. At my firm, the person who will architect your system is the person who writes your specification. That rule is non-negotiable.
Production Over Prototype, Every Time.
The AI industry has developed a pathological focus on demo-quality prototypes. “Look, our agent can do this in a conversation!” “Look, our model achieves 95% on this benchmark!” These are not useful statements for enterprise buyers. The question is: will this thing actually work in production? And if it does, will it keep working?
A shipped 85%-accurate model that runs reliably in production delivers more business value than a 95%-accurate prototype that never deploys. I scope engagements toward production outcomes. If an engagement cannot realistically ship to production within the scoped window, I say so before signing — not after. The firms that compete on prototype quality are optimizing for the wrong outcome. Production is the only outcome that matters.
Depth is a Defensible Advantage; Breadth is Not.
I founded a specialist firm specifically because I believe depth in one domain defeats breadth across many. A 20-person firm that does nothing but AI will beat a 200-person firm that does AI as one of ten services on engagement execution quality — because the 200-person firm has service-line politics, talent dilution, and attention fragmentation that the specialist firm does not.
This belief shapes hiring. We do not hire generalists who want to learn AI. We do not hire engineers transitioning into AI from adjacent disciplines. We hire engineers who have already shipped AI in production. It also shapes scope — we say no to adjacent service lines even when clients want us to add them. Each “no” strengthens the specialism. Over eight years, the compounding has mattered.
Responsible AI is Not a Policy Document.
Most firms address responsible AI by writing a policy page on their website. That approach satisfies procurement but does not change what actually gets built. Real responsible AI is enforced in the architectural decisions made during engagement scoping — not in an ethics statement published on a website.
At the firm, this means specific things: we do not train models on data we should not have had. We instrument every production model with explainability by default. We test for demographic bias as part of model evaluation. We design for human oversight on decisions that affect safety, credit, care, or employment. We refuse to build systems whose failure modes we cannot model. These are engineering decisions, not marketing decisions. They cost us specific engagements — clients who wanted us to build things we would not build. That is the right cost to pay.
If all five of those resonate, we will probably work well together. If two or three feel constraining, we are probably not the right match. I would rather know that early than find out during engagement delivery.
The ways to reach me directly.
Direct Contact
Email me
Real, monitored inbox. I read my own email. For engagement inquiries, response typically within 4 business hours during US, Europe, or Asia-Pacific business windows.
Book a 30-minute call with me
Free, 30 minutes, no pitch deck. I run these personally for most engagements. You can book directly on my calendar.
Send a detailed inquiry
If you want to describe your use case before scheduling, the main contact form goes to the same inbox as my direct email.
Elsewhere on the Web
Primary professional platform — regular posting on AI production patterns. Connection requests from enterprise roles accepted promptly.
Individual Top Rated Plus freelancer profile — shows client work history and verified engagement feedback.
Subscriber-based technical newsletter on production AI patterns. Approximately monthly cadence.
Light activity — occasional commentary on AI industry developments.
Long-form writing on technical AI topics.
Open-source contributions and public code.
Thirty minutes. With me, directly. No sales theater.
If you read this far and think we might be worth a conversation, the next step is a 30-minute call on my calendar. Free. No pitch deck. I run these myself for most engagements above a certain threshold — and I would rather tell you in that call if we are not the right partner than waste your time later.