Navin Kumar

AI Engineer building agent systems that don't fail quietly

About

I build LLM systems and the machinery that keeps them honest. That means agentic orchestration that runs unattended, retrieval that actually grounds an answer, and evaluation harnesses that tell you when a model has quietly gotten worse.

Most of my work starts from the same observation: ML systems rarely fail with a crash. They fail quietly. A recommendation drifts across sessions, a detector's per-class recall collapses after a retrain, an agent reasons confidently from a false premise. So I spend as much time on benchmarks, failure analysis, and supervision as I do on the models themselves. CSE at IIIT Vadodara, currently shipping agents in production.

Experience

AI Engineer

MAY 2025 –

DigitalxCode

Built an agentic stock intelligence platform on GPT-4o with a custom MCP orchestration layer. Four discrete tools (bucket generation, RAG retrieval, LLM health validation, DB write) autonomously produce investment signals across 24 equity buckets daily with no human in the loop. Added a ChromaDB + text-embedding-3-small memory pipeline so the agent reasons about portfolio drift across sessions instead of repeating itself.

Shipped it as a containerised 5-service system on GCP Compute Engine, running FastAPI, Next.js, PostgreSQL, Redis and ChromaDB via Docker Compose, with APScheduler handling nightly generation and comparison jobs behind a live dashboard.

Separately, fine-tuned an RT-DETR detector on VisDrone (10 classes) for real-time drone perception: 90%+ mAP across 5,000+ frames, latency cut to under 33 ms (30+ FPS) through optimization, batching, and runtime profiling. Built the benchmark suite that regressed model quality version over version, and ran failure analysis on misclassified samples to target the next retraining cycle.

  • GPT-4o
  • MCP
  • ChromaDB
  • RT-DETR
  • PyTorch
  • FastAPI
  • GCP

Software Development Engineer Intern

JAN 2025 – MAY 2025

EaseMyGrowth

Owned a time-series forecasting service end to end, covering exploration, model selection, Dockerized packaging, GCP deployment, and integration into a client-facing dashboard. Built reproducible pipelines over 10,000+ row datasets with Pandas and SQL, and integrated LLM workflows producing structured outputs for downstream BI consumers.

  • Time-series
  • Pandas
  • Docker
  • GCP

Software Development Engineer Intern

MAY 2024 – JUL 2024

EaseMyGrowth

Used BERT for sentiment analysis over platform reviews, classifying them positive / negative / neutral so the team could read customer feedback at volume instead of sampling it. Wrote a Python scraper to pull Amazon product reviews as an additional corpus for that analysis and for market research, and worked alongside computer-vision projects that shaped how I approach data processing today.

  • BERT
  • NLP
  • TensorFlow
  • Python
  • Docker
  • GCP

Independent Builder, Agent Systems

2025 –

Hackathons & open source

Four shipped agent systems across the OpenAI long-range-autonomy and SitRep Future of Work hackathons: a circuit breaker for long-running agents, a post-meeting research agent, a telemetry correlation engine, and a RAG résumé optimizer. Each was built end to end, from engine to evaluation to a deployed interface.

Projects

All on GitHub
Agent reliability Live

AgentFuse

A logical circuit breaker for long-range agents

Long-running agents don't crash. They drift. AgentFuse sits above the execution graph, watches the telemetry every framework already emits, and trips a breaker the moment a failure mode crosses a threshold. On a trip it freezes state, asks a separate reasoning model for a steering correction, and resumes. When recovery isn't safe it escalates to a human instead. One engine, three runtimes.

  • OpenAI AgentKit
  • LangGraph
  • Python
  • Observability
Dashboard Code
Observability Live

The Internet's Worst User Journey

Finding the users monitoring never sees

Errors are observable; frustration usually isn't. This reconstructs the path of customers who had a terrible experience while every dashboard stayed green, joining sessions, OpenTelemetry spans, and deploy diffs into one evidence-backed Ghost Story and a concrete fix. A deterministic correlation engine, not a model guessing at causes.

  • Node / Express
  • React
  • OpenTelemetry
  • SSE
Live demo Code
Retrieval Live

Relevant Section Identification

Which pages of this PDF answer my question?

Give it a PDF and a natural-language query and it returns the pages and section headings that answer it. Prose, tables and figures take three separate chunking paths so structure survives, then EmbeddingGemma vectors fuse with BM25 at 0.6. Evaluated on 45 graded queries with 12 held out: nDCG@5 0.645 to 0.728, hit@3 82% to 93%.

The interesting part is what did not work. HyDE, cross-encoder reranking and metadata scoring were each built and measured, and each lost ground. Every point came instead from repairing what the index contained: reading tables from the embedded text layer rather than OCRing a picture of them, and stripping the running header that made all 40 pages match a query equally. Retrieval quality here was never a ranking problem, it was an extraction problem wearing a ranking costume.

  • EmbeddingGemma
  • ChromaDB
  • BM25
  • unstructured
  • pdfplumber
  • Cloud Run
Live app Code
Autonomous research

Research Scout

A post-meeting research agent

Every meeting leaves loose ends nobody chases: open questions, competitors that got name-dropped, claims that went unchallenged. Research Scout mines the transcript for them, runs live web research on each in parallel, and returns one sourced briefing with inline links and honest caveats where sources disagreed.

  • Claude
  • Structured outputs
  • Web research
  • HMAC
Code
RAG

AI Résumé Optimizer

Retrieval that does real work

Embeds a résumé into a FAISS index, queries it with each requirement parsed from a job description, and flags anything with no semantically similar passage as NOT COVERED. The gap analysis is grounded in vector-search results rather than guesswork, and the rewrite is explicitly constrained never to invent experience.

  • FAISS
  • sentence-transformers
  • Flask
  • LLM scoring
Code
Computer vision

Human Activity Recognition in the Dark

LRCN for low-light video

A CNN + LSTM architecture for recognising activity in near-dark footage. Accuracy moved from 72% to 91% through targeted preprocessing and model tuning, validated against loss curves, validation accuracy, and confusion-matrix analysis rather than a single headline number.

  • PyTorch
  • CNN + LSTM
  • OpenCV
Code
Audio ML

Natural Sound Recognition

RNN audio classification

An RNN-based classifier over engineered Librosa features, taking accuracy from 70% to 90% through architecture tuning, with structured evaluation across loss, accuracy, and per-class confusion to show where the gains actually came from.

  • Librosa
  • RNN
  • TensorFlow
Code

Background

Education & Certifications

  • B.Tech in Computer Science & Engineering (2021–2025)
  • Indian Institute of Information Technology (IIIT) Vadodara
  • Machine Learning Specialization, DeepLearning.AI and Stanford (Sep 2023). Covers supervised learning, advanced learning algorithms, and unsupervised learning with recommenders and RL. Verify
  • Introduction to Generative AI, Google Cloud (Aug 2024). Verify
  • Fundamentals of Deep Learning, NVIDIA Deep Learning Institute

Retrieval Strategies

  • Vector search: dense cosine over EmbeddingGemma-300m vectors in ChromaDB
  • Keyword search: BM25 across the same chunk set
  • Hybrid search: the two fused at alpha 0.6 over min-max normalised scores, chosen from a sweep showing a plateau across 0.5 to 0.8
  • Section expansion: a chunk inherits a decayed score from the best chunk under its heading, capped at four to stop one weak hit injecting a whole section
  • Page aggregation: chunk hits collapse to pages within 55% of the top score, capped at eight, since a fixed top-k is wrong in both directions
  • Metadata per chunk: page number, section heading, chunk type, table source
  • Measured and rejected on this corpus: HyDE with Qwen2.5-1.5B-Instruct (nDCG@5 0.710 to 0.684), cross-encoder reranking with BAAI/bge-reranker-base (0.728 to 0.727), and metadata as a separate scoring channel (0.710 to 0.704)

Document Processing

  • Layout-aware partitioning with unstructured hi_res, typing every block as title, prose, table or image
  • Three chunking paths that never share code, so structure survives: prose grouped by section title, tables to markdown, figures to caption plus OCR
  • Chunk sizing: 1000 minimum, 3000 soft, 5000 hard, with 500 characters of overlap applied between adjacent chunks and not only inside oversized ones
  • Tables read with pdfplumber from the embedded text layer rather than OCR of a rendered image, which recovered a symbol column the index was missing entirely
  • Generic boilerplate detection: lines appearing on 60% or more of pages are dropped, which cleared 226 of 852 elements and 57 false section titles
  • Tesseract OCR with stopword filtering for figure keywords

Models & Agents

  • Supervised fine-tuning (SFT): GPT-4o, BERT, RT-DETR
  • Agent frameworks: LangGraph, OpenAI AgentKit, LangChain, MCP orchestration
  • Embeddings: google/embeddinggemma-300m, an asymmetric prompt-conditioned model at 768 dimensions, plus text-embedding-3-small
  • Cross-encoder: BAAI/bge-reranker-base
  • Generation: Qwen2.5-1.5B-Instruct for grounded answers and HyDE passages
  • Vision: llava-interleave-qwen-0.5b for figure captioning, chosen over llava-1.5-7b because 4 GB of VRAM could not hold the larger model
  • Vector stores: ChromaDB with cosine HNSW, FAISS

Evaluation & Reliability

  • Retrieval metrics: nDCG@k, hit@k, recall@k over graded relevance, where a complete answer and a passing mention score differently
  • Held-out splits: 12 of 45 queries never inspected while tuning anything
  • Failure taxonomy: queries tagged by kind (word, symbol, table, figure, multi) because those classes fail for different reasons and an average hides it
  • Detection metrics: mAP, per-class precision/recall, latency
  • Version-over-version quality regression across index builds
  • Runtime profiling: separating what the GPU helps (indexing, 7x) from what it does not (query time, 117 ms against 119 ms)

Technical Expertise

  • Languages: Python, C/C++, JavaScript, SQL
  • ML: PyTorch, TensorFlow, Keras, Hugging Face, scikit-learn, OpenCV, Librosa
  • Data: Pandas, SQL, reproducible data pipelines
  • Production: FastAPI, Next.js, PostgreSQL, Redis, Docker, GCP
  • CI/CD: GitHub Actions, automated build and deploy pipelines
  • Familiar with AWS (ECS/EKS), Git/GitHub, REST APIs

Writing

Why agents fail quietly

The four long-horizon failure modes (loops, drift, logical traps, and silent budget burn), and why the agent is the worst possible judge of whether it hit one.

Draft

RAG that actually retrieves

Most RAG demos would work identically with the retrieval removed. A test for whether yours is doing real work.

Draft

Shipping a detector under 33 ms

What actually moved the needle fine-tuning RT-DETR for real-time drone perception, and what turned out to be noise.

Draft

Let's build something reliable.

I'm open to AI Engineer roles and interesting problems in agent orchestration, retrieval, and model evaluation. The inbox is genuinely open.

navin.kumarr0437@gmail.com LinkedIn GitHub