saarthak
building things I assumed were impossible
I'm Saarthak, an ML engineer who builds end-to-end ML systems at startups — from autonomous knowledge graph agents to novel GNN models trained from scratch. Check out my posts section :)
what i do
Applied AI Intern • Brace.so
Mar 2026 - May 2026
- Pedro (Autonomous Knowledge Graph Agent): Built an autonomous agent that maintains a 10K+ node knowledge graph without manual curation, using a multi-LLM architecture (Gemini + Anthropic Claude) with parallel evaluation and a deferred decision system for low-confidence candidates. The agent extracts skills and tools from profile text, ranks evidence with semantic similarity and usage patterns, and classifies terms into a parent-child taxonomy — all without human intervention.
- Search Engine: Rebuilt people search ranking with LLM-generated structured field constraints (xAI/Grok), replacing fuzzy matching with deterministic gating for exact role, company, and location requirements.
- Oden (Connection Path Discovery): Practical implementation of six degrees of separation given any target person, finds the exact chain of intermediary people connecting you to them. Built multi-hop pathfinding across 4,739 actors and 11,609 edges, evaluating 3.4M+ directed trials and 29.3M possible route combinations. Achieved 99.87% path-finding success rate across 9 strategies. Product output directly attracted early-stage VC interest.
- Actor-Relationship Optimization: Ran 60+ controlled experiments to design a relationship-scoring strategy that reduced processing to 2.93% of baseline while retaining 98% of close-connection edges, across 72 profiles and ~90K baseline relationship records.
- Knowledge Graph Visualizer: Built a full-stack knowledge graph visualizer for exploring and debugging the graph interactively.
ML Engineer • Stealth Startup (Contract)
Jun 2025 - Nov 2025 (Contract work: Jul 2025 - Oct 2025)
- Novel Everything: Given a $35 budget to solve a problem nobody had solved : trained a GNN from scratch for molecular property prediction with a novel pharmachemistry-task, novel dataset, and novel model. No pretrained weights, no reference implementations, no existing benchmarks.
- Experimentation at Scale: Adapted and extended GraphGym to run experiments across 48k data points. Orchestrated 1,000,000+ optimization trials on remote distributed GPUs and beat Chemprop, the industry standard.
- Delivery: Shipped a production-ready, lightweight (2M param) inference API in 8 weeks, handling the full lifecycle from data ingestion to AWS deployment using FastAPI and Docker.
- Business Impact: Client used this work to secure a ₹30 Lac (~$35k) research grant.
Software Engineering Intern • HeydoTech
Jan 2024 - May 2024
- Utilized Python libraries including OpenCV and PyAutoGUI to develop 2 automation tools for internal projects
- Collaborated with cross-functional teams of 5+ members to implement feature enhancements, improving system performance for approximately 100 users
- Contributed to UI/UX improvements that enhanced overall user experience, as measured by positive feedback from 80% of test users
projects
Docbook
- GraphRAG-based CLI tool for intelligent documentation context extraction
- Single URL input to full knowledge graph conversion using crawl4ai
- Automated semantic search and retrieval system
Bayesian Course Recommendation System
- Achieved 0.915 AUC using Bayesian Personalized Ranking for implicit feedback
- Implemented matrix factorization with BPR optimization
- Handles cold-start problem with probabilistic inference
education
BTech, ECE
IIIT Delhi
Nov 2022 - Present
what excites me
I'm drawn to problems that feel impossible at first — autonomous agents that maintain knowledge graphs without human curation, search engines that understand structured intent, GNN models trained from scratch on datasets that didn't exist before.
Right now I'm deep in NLP and graph ML, building systems that ship to production and solve real problems at startups.
let's connect
Want to discuss AI, collaborate on projects, or just chat about tech? Feel free to reach out!
posts
Graph Neural Networks from Scratch
A comprehensive exploration of Graph Neural Networks — from ground up. Covering everything from how to represent Graph to how contraction mapping gurantees convergence in training of GNN.
Read post →Bayesian Personalized Ranking
Exploring the fundamentals of Bayesian Personalized Ranking for implicit feedback systems. Breaking down the math, intuition, and implementation details behind one of the most elegant recommendation algorithms.
Read post →Bayesian Linear Regression
A deep dive into Bayesian approaches to linear regression. Understanding posterior distributions, conjugate priors, and how Bayesian inference provides uncertainty quantification unlike traditional least squares methods.
Read post →