Research in. Production out.
I'm Sujit. I build at the point where machine learning stops being a research artifact and starts becoming a real system.
My work sits across Machine Learning, Deep Learning, LLMs, agentic systems, retrieval, evaluation, computer vision, and product engineering. I care about what happens after the demo: messy inputs, imperfect models, latency, observability, cost, reliability, and the engineer who eventually has to operate the thing.
I've worked on systems that moved revenue, user engagement, and engineering velocity — not just benchmark scores.
The loop I enjoy:
BUILD → MEASURE → BREAK → LEARN → SHIP
The interesting engineering starts when the happy path stops being representative.
Numbers I can defend in an architecture review, not vanity metrics:
| Where | What I did | Measured Result |
|---|---|---|
| Launchpad.AI | Architected agentic RAG with Neo4j + FAISS + LangGraph; ran LLMOps A/B tests across retrieval & ranking strategies | +2% revenue |
| Endure Solutions | Shipped a full-stack 3D visualization platform; iterated on UX using real usage telemetry | −23% bounce rate |
| Endure Solutions | Automated code review + CI/CD workflows across the engineering team | +27% developer velocity |
| Data-Thon 2025 | Built a scalable ML system for e-commerce item-accuracy under a 48-hour constraint | 🥈 2nd place |
| Hackspiration · Devfolio | Shipped an AI/ML solution under a national hackathon deadline | Top 30 nationwide |
| HeadOn | Designed a structured AI training-data collection pipeline across a distributed team | 30+ contributors |
Every claim above traces to an eval, an A/B test, a measurable product metric, or a delivery outcome. Ask me for the methodology — I have it.
The problems I keep coming back to when nobody's assigning me tickets. Most of my next projects start somewhere in this section.
| Track | What I'm actually doing | What I'm chasing |
|---|---|---|
| Local LLMs from scratch | Implementing a coding-focused LLM from first principles — tokenizer, embeddings, attention, training loop, and local inference | Owning every layer instead of hiding the interesting parts behind an API; building toward local-first, private coding agents |
| LLMs as research instruments | Using local models for literature triage, hypothesis stress-testing, mathematical sanity checks, and evidence extraction | The gap between an agent that accelerates research and one that confidently invents evidence — closing it with evaluation |
| Dark matter | Working through galaxy rotation curves, gravitational lensing, CMB power spectra, structure formation, and particle candidates | A theory supported by overwhelming indirect evidence, yet still lacking a confirmed direct detection. That tension is the interesting part |
| Black hole theory | Exploring Hawking radiation, the information paradox, Page curves, island formulas, and singularity theorems | What happens at the boundary where general relativity and quantum mechanics refuse to fit together cleanly |
Fair warning: if my commit graph goes quiet for a week, I'm probably stuck inside a Penrose diagram. I'll be back — usually with a repo.
- 🕸️ Agentic RAG & knowledge graphs — multi-hop, entity-aware retrieval; the hard question isn't "can the agent answer?" but "does it know when it shouldn't?"
- 🛡️ LLM evaluation as an engineering discipline — hallucination detection, jailbreak testing, trust scoring, regression suites, and eval harnesses that catch failures before users do
- 👁️ Applied computer vision — inference pipelines designed for real hardware, real latency, and real failure modes
- ⚙️ Production discipline — observability, CI/CD, containerization, rollback, and recovery; an unmonitored model isn't shipped, it's a liability with an API key
- 💸 Inference economics — cost-per-query is a design constraint, not an afterthought; caching, routing, quantization, distillation, and efficient serving often beat simply using a bigger model
Skeptical of: agent frameworks with more YAML than reasoning, benchmark numbers nobody has tried to break, and "evals" that are three vibe checks in a trench coat.
01. Design before implementation
Two paragraphs of architecture can save two weeks of refactoring.
02. Every model ships with an eval.
Every eval ships with a failure taxonomy.
03. If you can't measure it, you don't understand it.
If you can't observe it, you can't operate it.
04. Production is part of the design.
Latency, reliability, cost, rollback, and recovery are model features.
05. Optimize the system before optimizing the model.
06. Complexity has to earn its place.
07. Delete more code than you write.
| Layer | Stack |
|---|---|
| Core | Python · Tensorflow · PyTorch · C++ · FastAPI · PostgreSQL |
| AI / Data | Agentic RAG · Knowledge Graphs · LangGraph · FAISS · OpenCV · LLM Evaluation |
| Infrastructure | Docker · Kubernetes · Redis · GCP · AWS · GitHub Actions · CI/CD |
| Currently Exploring | Quantization · KV-cache optimization · Inference serving · Distributed inference |
Because eventually every AI engineer discovers the same equation:
tokens × latency × throughput × memory × hardware × traffic = the real bill
I write about the research → production seam on Hashnode — architecture decisions, engineering trade-offs, failure post-mortems, and evaluation methodology.
If you want to understand how I think, the writing is probably a better signal than a list of technologies.
- Agentic AI, production ML, AI infrastructure, or research engineering
- Systems where distributed systems + ML actually have to work together
- Technical collaboration, architecture reviews, and projects with difficult failure modes
- Problems where the problem is harder than the framework
I like building from 0 → 1, especially when nobody has figured out the right abstraction yet.
"In God we trust, all others must bring data." — W. Edwards Deming And if you bring data, bring evals.