IbrahimSultan.

An answer a lawyer can use has to point back to the law it came from.

A legal research assistant is only useful if every claim leads back to the law it came from. I hardened the Python and FastAPI backend that decides where to look, gathers the evidence and attaches the citations, for legal research and for a lawyer's own uploaded documents.

~10,000

Indian acts searched; every answer cites the one it relied on.

My contribution

• Routing: each question goes to internal Vespa legal search, a bounded web search, or a request for clarification.
• Contextual retrieval: Reciprocal Rank Fusion across several LLM-generated queries, over chunks enriched with their document summaries.
• Gemini fallback through LiteLLM before the first model output, with cancellation handling and sanitized streaming errors that keep partial answers.
• Per-user token budgets admitted with renewable Redis locks, so one user cannot exhaust shared capacity.
• Subscription checks on uploads, downloads and files carried through chat history.
• AI web tools restricted to validated public HTTPS URLs and bounded searches; request logging that keeps private queries, responses and identifiers out of routine logs.

Python · FastAPI · Vespa · LiteLLM · Redis

Over 900,000 judgments and acts, one search box, and a migration that lost nothing.

One search across Supreme Court and High Court judgments, central and state acts. I built the search engine behind it and moved the whole index onto production hardware.

900,000+

legal documents indexed and searchable.

~100×

faster embedding and indexing on a GPU server.

My contribution

• Automated ingestion and vectorization of 900,000+ Supreme Court, High Court and statutory acts from AWS S3.
• Provisioned a Vast.ai GPU server with CUDA 12 and PyTorch, making nomic embeddings about 100 times faster.
• Migrated the full index from the GPU machine to a public Hetzner CPU deployment with S3 volume backups and no data loss.
• Hybrid retrieval: BM25 keyword search for exact legal terms and HNSW vector search for meaning, merged with Reciprocal Rank Fusion; Vespa schema and indexing adapted for CPU query performance.
• Enterprise connectors (S3, Dropbox, Google Drive, MinIO) with role-based access control.

Vespa · Docker · AWS S3 · CUDA 12

A 300K-token brief, drafted one section at a time, with the lawyer holding the pen.

Long legal drafts, written section by section with a lawyer in control: they set the structure first, then the system writes one section at a time and carries context between them.

300K

tokens per upload, enough for long briefs and contracts with annexures.

My contribution

• Lawyers review, edit and reorder the title and outline before any text is generated.
• Section-by-section generation with streamed progress, ending in a final PDF with a persistent short link.
• Rolling context: each finished section is summarised and carried forward, so later sections stay consistent within token bounds.
• Took over the CLM Auto Drafting front end to ship it on time, and kept in-progress drafts intact across a hard refresh.
• Built and containerized local document-partitioning and Granite embedding services with batched inference and vector-dimension validation.

FastAPI · Gemini · React · Next.js · Docker

Four collections, four shapes of data, one box to search them all.

Reworked search across Supreme Court judgments, High Court judgments, central acts and state acts: four collections with different shapes, searched from one box.

Separate MongoDB aggregation paths per court and act scope, compound indexes, timeouts on every query, and tolerant matching for judge names and case titles (v, vs, v.). Dates and legal entities are normalized, and the React interface gained mixed-source pagination, filters, highlights and search history.

MongoDB · Python · TypeScript · React

A whole GenAI product: search, chat, outcome analysis and the servers underneath.

At a small product company I led the architecture and built across the whole stack of a public AI platform for legal users: retrieval-augmented chat, case-outcome analysis, document generation with a human in the loop, data ingestion and the deployment underneath. Outcome scores come from a local classification model with structured reasoning; they are model outputs, not guaranteed outcomes.

40%

less time spent on legal research.

$20/mo

flat plan replaced about $375 an hour of usage-based AI spend.

100+

tribunal orders ingested every day, behind case-outcome dashboards.

My contribution

• Retrieval-augmented chat on Vespa (Sentence Transformers, keyword and vector search, reranking), with streaming responses and OCR support.
• Seven providers behind LiteLLM (Gemini, Groq, Cerebras, Hugging Face, Ollama, Claude and OpenAI) for flexible model routing.
• NLP layers for intent detection, classification and reranking, and GPU-based asynchronous model servers for inference and embeddings.
• Automated ingestion of tribunal orders with Python, SQL and OCR, structured into datasets for case-outcome dashboards.
• Enterprise search with 40+ connectors (Slack, Google Drive, GitHub and more), multi-tenant workspaces, RBAC, OAuth2 and SSO, deployed with Docker Compose, Redis, NGINX and Cloudflare Tunnels.
• Mentored developers on system design, API development and deployment practice.

FastAPI · Svelte · Vespa · LiteLLM · PostgreSQL

Some legal PDFs are half typed, half scanned. The old pipeline only read half.

Re-engineered the extraction pipeline to read hybrid PDFs page by page, fixing a legacy bug where native text let scanned pages skip recognition. Only scanned pages go through OCR, in English, Hindi and Marathi.

Encrypted PDFs are supported, and one extraction interface covers PDF, DOCX, PPTX, XLSX, EML, EPUB and HTML, with temporary-file cleanup for large legal documents.

OCRmyPDF · Tesseract · MarkItDown · Python

An agent is a loop. Making the loop trustworthy is the engineering.

Behind the research assistant I engineered the runtime itself rather than relying on framework defaults: a custom agent loop that plans, calls tools in parallel where it is safe, streams its answer and keeps the model’s reasoning out of what the user reads. Underneath, models are served in explicit tiers, so a workflow can move between hosted and self-hosted models without being rewritten.

My contribution

• A custom agent loop for chat state, tool orchestration, retries and streamed output, using LangChain core components only where they earn their place.
• Independent tool calls (retrieval, search, document processing) run in parallel and merge into one synthesis, cutting latency on multi-step research.
• A streaming parser that separates reasoning markers such as <think> from the final answer, so internal traces never reach the screen.
• Deep Research orchestration: multi-step planning, autonomous tool calling, web search and evidence retrieval before the answer is written.
• Model tiers behind LiteLLM, and a serving path moved from Hugging Face Inference to Google’s Generative AI SDK while staying provider-agnostic.
• Containerized local document-partitioning and Granite embedding services: batched normalized inference, bounded request queues, thread offloading and vector-dimension validation.

Python · LiteLLM · LangChain · Agents · Embeddings · Docker

Credential-stealing releases of two popular packages hit our stack. I hunted them down.

When credential-stealing releases of litellm (v1.82.7/1.82.8) and axios (v1.14.1/0.30.4) appeared, I found them in our stack and removed them. I wrote the organization-wide remediation guide, hardened dependency lockfiles and moved build pipelines to secure baselines.

Incident Response · Security Audit · LiteLLM · Axios · Dependency Hardening

Teaching a cancer-risk model to listen to the clinicians who use it.

IXCAN-A predicts colon-cancer risk to support clinical decisions. My part was the data and the feedback loop: cleaning and preparing training data, validating behaviour against expected clinical outcomes, and turning clinicians' feedback into model refinements.

92%

accuracy reached by the model I helped refine.

2

hospitals adopted the system.

A contribution to a team's model. Not a medical guarantee.

Python · Machine learning · Data preparation · Clinical feedback