AI Engineer · Backend · Formula 1 fan

Building production LLM systems: RAG pipelines, AI agents, LLM evals, inference gateways and FastAPI backends

Python and FastAPI systems with embeddings, vector databases and tool‑using agents — tuned like a race car: measured, iterated, and judged on the data.

See the work Visit the pit wall

Timing tower

Measured, not guessed.

Every system ships with a number.

  1. P1MedRAG Top‑1 retrieval accuracy · PubMedQA 0%
  2. P2Strategy Agent Eval pass rate · gpt‑oss‑120b, 39 cases 0%
  3. P3Backtester Parameter combinations in under 1s · Numba 0
  4. P4User Hub Products made reachable across client tenants 0+
  5. P5User Hub Less manual configuration time 0%
  6. P6ITIS Faster issue resolution 0%

S1 Experience

Zycus Infotech.

AI Engineer Intern · Mumbai, India · Jun 2026 – Present

Access gateway

User Hub

Internal access gateway giving employees secure access to 100+ products across client tenants. User, client and project mappings grant access automatically.

0%less manual configuration time
  • Python
  • FastAPI
  • OAuth 2.0
Teams chatbot

Zyra

IT self‑service chatbot in Microsoft Teams automating 8 Microsoft 365 admin workflows — mailboxes, distribution lists, and end‑to‑end onboarding and offboarding.

  • Copilot Studio
  • Power Automate
  • Entra ID
AI agent

Natural‑language query layer

A LangGraph agent with MCP tools and structured prompts over tenant, product and workforce services, replacing manual navigation across internal systems.

  • LangGraph
  • MCP
  • Prompt engineering
Issue resolution

ITIS

FastAPI backend and Docker deployment for an internal issue‑resolution portal, automating ticket routing.

0%faster resolution
  • FastAPI
  • Docker

S2 Projects

Built end to end.

Agents, retrieval and inference infrastructure — each with its own evaluation.

LLM agent · Evals

LLM Strategy Agent with Eval Harness

A tool‑using agent that turns plain‑English trading ideas into strategy code — with a hidden‑data eval harness and a fast backtesting engine underneath.

0%gpt‑oss‑120b pass rate
0eval cases
<1s25k param combos
Details
  • Iteratively engineered prompts, static checks, a subprocess sandbox, a decline tool for impossible requests, and rate‑limit‑aware failover across providers.
  • Eval harness grades code on hidden data against reference implementations and detects look‑ahead by perturbing future bars; gpt‑oss‑120b passed 92% vs. 28% for local Qwen2.5‑Coder 7B.
  • Event‑driven backtesting engine with a Numba fast path and GitHub Actions CI that verifies the eval grader.
  • Python
  • FastAPI
  • React
  • Numba
  • Ollama
GitHub
RAG · Healthcare

MedRAG — Medical Document QA

Full‑stack RAG that answers medical questions grounded only in uploaded PDF, DOCX and TXT files, with streamed answers and cited sources.

0%top‑1 retrieval
0%expert agreement (human 78%)
0smedian latency
Details
  • Document preprocessing and chunking, pgvector (HNSW) search, and Tesseract OCR.
  • Evaluated on PubMedQA (1,000 expert‑labelled questions): 98.0% top‑1 and 99.4% top‑4 retrieval accuracy.
  • Gemini‑to‑Groq LLM fallback chain and a tuned similarity cutoff that rejects off‑topic queries before any LLM call, wrongly refusing only 2 of 900 valid questions.
  • Python
  • FastAPI
  • LangChain
  • pgvector
  • React
  • Docker
GitHub
Thesis · In progress

Adaptive LLM Inference Gateway

An OpenAI‑compatible gateway for adaptive cost‑quality‑latency routing across open‑weight LLM fleets. Race a request through it ›

2‑tierexact + semantic cache
SHA‑256Redis exact match
LinUCBrouter (in progress)
Details
  • Request normalization so equivalent requests share one cache key.
  • Semantic cache with BGE‑small embeddings and FAISS, with similarity scores calibrated into hit probabilities via isotonic regression.
  • In progress: vLLM‑served Qwen2.5/Llama replicas, a GPU‑time cost model, circuit breakers, and a LinUCB router.
  • Python
  • FastAPI
  • Redis
  • FAISS
  • Docker
GitHub

S3 Pit wall

Race your request.

A playable sketch of my Adaptive LLM Inference Gateway. Every request is a car — cache hits take the pit‑lane shortcut, misses run the full lap through a model.

0requests
0%cache hit rate
Exact cacheRedis · SHA‑256 key
Semantic cacheFAISS · BGE‑small
Full inferenceRouter → LLM replica
Timingsim
  1. Waiting for lights out…

Simulated latencies for illustration only — not benchmark results.

aditya@pitwall: ~

Garage

The toolkit.

Drag to spin

Languages

  • Python
  • SQL
  • C++
  • TypeScript

AI/ML & LLM

  • LangChain
  • LangGraph
  • sentence‑transformers
  • Ollama
  • MCP
  • OpenAI
  • Anthropic
  • Gemini
  • Groq

Generative AI & NLP

  • Prompt engineering
  • RAG
  • Embeddings
  • Vector databases
  • NLP
  • AI agents
  • LLM evaluation
  • Semantic caching
  • LLM routing & fallback

Microsoft 365 & Automation

  • Copilot Studio
  • Power Automate
  • Microsoft Teams
  • Entra ID
  • Exchange admin

Backend & Frameworks

  • FastAPI
  • REST APIs
  • asyncio
  • Pydantic
  • SQLAlchemy
  • Pandas
  • Numba
  • React
  • pytest

Databases & Infrastructure

  • PostgreSQL (pgvector)
  • Redis
  • ChromaDB
  • FAISS
  • Docker
  • AWS S3
  • Git
  • GitHub Actions

Education

Vellore Institute of Technology, Bhopal.

Integrated M.Tech in Artificial Intelligence · 2022 – 2027 (Expected) · CGPA 7.76

Box, box.

Let’s build
something fast.

Open to AI engineering and backend roles.

Email me