Available for Fractional AI Advisory & Architecture Sprints (Q4 2026)
I'm an ML engineer who has done the use case, a founding engineer who takes it from ideation to POC to MVP, and the person who runs the experiments, talks to stakeholders in terms of burn, ARR, and the user, and designs the system and the infra from top to bottom.
The work is the path from a rough model to something a user will trust: broadcast vision under 15 ms, FinOps agents that took the manual work out, document pipelines that only bother a person on the uncertain fields, cited answers, bounded agents, voice, code, video, fine-tunes, and support that knows when to stop. I run the experiment, choose the architecture and the infra, and say what it does to burn, the quote, or the ticket.
Modeling · Experiments · Ideation to MVP · Architecture and infra · Stakeholders, burn, and the user
Small upstream fixes in PyTorch and YOLO-World.
New Delhi · AI/ML
About
I have shipped the systems on the projects page. Live sports vision that had to stay under the broadcast ceiling. FinOps agents and a retrieval gateway where a wrong answer or an open loop showed up as review time and GPU spend. Document extraction that only queued the uncertain fields. Then the later work: cited answers, agents with a step cap, voice that will not write a field early, a coding agent that cannot merge itself, video that does not send every hour to a cloud GPU, fine-tunes that have to beat the base model, and support that counts a reopen as a failure.
Each of those started as an experiment, became a POC, and only shipped once the architecture was decided: what the model may emit, where it runs, and what a wrong output costs the user. I set that expectation with the people paying for it. Inference cost and token burn are part of the design, not a finance note after launch. BTech in Computer Science (AI/ML). A few small upstream fixes in PyTorch and YOLO-World.
2024 — Present
Senior ML Engineer & Architect · Independent
End-to-end product ownership, FinOps agents (90%+ less manual handling), org-wide LiteLLM, stakeholder-facing roadmaps.
2023 — 2024
Computer Vision Engineer · Document intelligence · B2B
Supply-chain vision platform, enterprise RAG on Bedrock/Qdrant, production MLOps and team mentoring.
2022 — 2023
AI Engineer · Sports broadcast · Startup
First real-time AI broadcast product in production—TensorRT pipelines at 200 FPS and sub-15 ms delivery.
What I bring
Founders hire me when they need a fractional co-founder/CTO for AI: someone who can shape the use case, ship the system, and keep stakeholders aligned while burn stays honest.
01 · Product
From problem framing and MVP scope to MLOps, observability, and the launch checklist. I own the thread so you are not juggling three vendors and an internal team that never met.
Use-case selection · Roadmap · POC → pilot → GA · SLOs & on-call
02 · Stakeholders
Clear narratives for founders, investors, and customers—what the model will and will not do, when humans stay in the loop, and how success is measured before spend ramps.
Executive readouts · Demo vs production gaps · Risk registers
03 · Business
Burn-aware sequencing: which features earn their GPU bill, which automations replace headcount, and which experiments should die in week two. I optimize for runway and defensibility, not vanity benchmarks.
Unit economics · Build vs buy · Compliance as a ship gate
04 · ML depth
Multi-agent orchestration with bounded steps, hybrid retrieval, fine-tuning and evals, and serving on Triton or vLLM with cost-per-query tracing. Plus vision and multimodal when the product demands it.
Agents · RAG · SFT/RLHF · LiteLLM · Qdrant · vLLM · TensorRT · CV
05 · Leadership
Hiring bar, code review culture, and hands-on mentoring so your engineers inherit patterns—not dependency. I have led pods through production incidents and compliance reviews.
Tech lead · Hiring loops · Runbooks · Junior → mid growth
Latency
<15 ms
End-to-end real-time edge CV inference for global live sports broadcast feeds.
Hallucination cut
30%
Achieved across enterprise RAG and agentic workflows via hybrid BM25 + dense vector routing.
OCR accuracy
+25%
Layout-aware document extraction pipelines using LayoutLMv3 and deep learning inpainting.
FinOps automation
90%+
Manual invoice and payment handling removed via autonomous agents on a regulated B2B FinOps platform.
ML & platform depth
Agentic workflows, retrieval, inference serving, and vision—architected with cost, compliance, and operating metrics from day one.
01
Customer support copilots, internal ops agents, and document-to-action pipelines—built with step limits, tool routing, and eval harnesses so loops cannot silently burn budget or bypass policy.
LangGraph · CrewAI · Smolagents · LiteLLM · guardrails
02
Hybrid BM25 + dense retrieval, reranking, and citation-aware answers over PDFs, tickets, APIs, and SQL. The patterns enterprises ask for: access control, freshness, and hallucination metrics in CI.
Qdrant · Chroma · Pinecone · LlamaIndex · Bedrock · RAGAS
03
Org-wide gateways, model routing by task, TTFT and latency SLAs, and Prometheus/Grafana dashboards that tie every feature flag to dollars per query.
vLLM · Triton · LiteLLM · Ollama · LoRA/QLoRA · SFT
04
Layout-aware document AI, real-time edge inference, and redaction at scale—when the product is pixels and milliseconds, not chat alone.
PyTorch · TensorRT · LayoutLM · OpenCV · SageMaker · ECS
Case studies
Systems I have taken from a first model to something a user and a stakeholder will trust.
Advisory models
Two modes: a fixed sprint to decide what to build and how, and a monthly retainer when you want a fractional co-founder/CTO for AI—product, architecture, and team in one seat.
2-week fixed scope
$6,000 – $8,000
Fixed fee · 2 weeks
Seed and Series A teams deciding what to build, what to defer, and whether the current stack will survive first real users.
Monthly retainer
$10,000 – $14,000 / month
Retainer · 15–20 hours/week
Funded teams that need a senior hands-on technical lead without a six-month hiring cycle or equity dilution.
Schedule
Tell me what you are shipping and where it hurts—latency, hallucinations, compliance, or cost. We will map the architecture in a working session, not a sales call.
Contact me