Manas Vardhan

AI Developer & Researcher — Los Angeles

scroll
01
01

I build intelligent systems at the intersection of research and product — from multi-agent architectures to production ML pipelines. My work spans founding Friday Inc, engineering at scale for JPMorgan Chase, and publishing research on temporal stability in latent spaces. I'm drawn to problems where the math is beautiful and the impact is real.

"I'd rather build one system that actually works than ten that almost do."
Languages Python, SQL, JavaScript
Frameworks PyTorch, TensorFlow, Transformers, FastAPI, LangChain
Domains LLMs, Computer Vision, NLP, Multi-Agent Systems, MLOps, RAG
02
02

Experience

USC
2025 — Present

Graduate Researcher

Computational Physics Group, USC

Built a novel sparse-view 3D reconstruction pipeline using multi-view latent fusion on frozen vision backbones, cutting compute requirements while matching fine-tuned model quality. Leading 6 independent research projects targeting NeurIPS and EMNLP 2026 across LLM agents, reasoning efficiency, prompt robustness, and video generation.

USC GSG
Aug 2025 — Present

Finance Intern

Graduate Student Government, USC

Managing a six-figure annual budget allocation across 200+ graduate student organizations. Streamlining reimbursement workflows and financial reporting processes, reducing processing time by 40%. Coordinating funding proposals and expenditure tracking for campus-wide graduate initiatives.

Friday Inc
2025 — Present

Founder & AI Developer

Friday Inc

Building multi-agent systems and AI-native developer tools from zero to one. Shipped 6 open-source Python libraries for the LLM ecosystem covering cost tracking, prompt versioning, safety validation, benchmarking, and agent observability. Active open-source contributor to HuggingFace Transformers, scikit-learn, PyTorch, and MLflow.

JPMorgan
2023 — 2025

Software Development Engineer

JPMorgan Chase

Enterprise-scale ML and data science engineering across trading and risk platforms. Built production pipelines processing millions of daily transactions with real-time anomaly detection and risk scoring.

SimpliClarify
2021 — 2022

Data Scientist

SimpliClarify

Built NLP pipelines for document intelligence products, automating extraction and classification across unstructured enterprise data.

SmokeTrees
2020 — 2021

ML Engineer

SmokeTrees

Shipped computer vision and deep learning models to production, from prototype to deployment at scale.

03
03

Education

USC
2025 — 2026

M.S. Computer Science

University of Southern California

Research in 3D generation using large reconstruction models, modelling physical systems with transformers, and latent-space temporal stability.

VIT
2019 — 2023

B.Tech Computer Science, Minor in Business Systems

Vellore Institute of Technology

Undergraduate thesis on generative modelling with diffusion-based architectures for high-fidelity image synthesis. Coursework in algorithms, distributed systems, and statistical machine learning.

04
04

Research

9 research projects spanning 3D generation, LLM agents, reasoning, and robustness. Targeting NeurIPS, EMNLP, and NAACL 2026.

2026 Prompt Sensitivity Landscapes — Discovered that synonym substitutions destabilize LLM outputs more than typos or deletions. Identified prompt structures that are 3.6x more robust to perturbation. Targeting EMNLP 2026
2026 Tool Use Reliability in LLM Agents — Built TUR-Bench (30 cases, 6 failure types). Found that LLMs skip available tools 36.7% of the time but never hallucinate fake tools. Failed responses are 5x longer than correct ones. Targeting EMNLP/NAACL 2026
2026 Agent Memory Architectures — Designed CAM (Consolidation-Augmented Memory), a biologically-inspired memory system with episodic buffering, semantic consolidation, temporal decay, and interference resolution. Targeting NeurIPS 2026
2026 Adaptive Test-Time Compute Allocation — Training RL policies (PPO) that dynamically allocate reasoning compute per step based on task difficulty, targeting 40% compute reduction with less than 2% accuracy loss. Targeting NeurIPS 2026
2026 Efficient Reasoning Distillation — Distilling verbose chain-of-thought traces from large models into compact reasoning for smaller models, with NLI-based faithfulness scoring to verify logical preservation. Targeting NeurIPS 2026
2026 Latent-Space Temporal Stabilizers for Video Diffusion — Lightweight temporal consistency modules that plug into existing video diffusion models to reduce inter-frame flickering without retraining. Targeting NeurIPS 2026
2025 Multi-View Latent Fusion on Frozen Backbones for Sparse-View 3D Reconstruction — Reconstructs 3D objects from as few as 4 views by reusing frozen pre-trained vision encoders, eliminating costly end-to-end fine-tuning while matching fully-trained baselines
2024 LLM Reliability in Production Systems — Catalogued hallucination patterns, latency spikes, and degradation under load. Developed monitoring heuristics for detecting reliability regressions before end-user impact
2023 Bias in Automated Credit Scoring — Audited ML credit scoring pipelines for demographic bias across race, gender, and age. Proposed mitigation strategies that reduced bias while maintaining predictive accuracy
05
05

Projects

LLM Cost Guardian

LLM Cost Guardian

Real-time cost tracking for LLM API calls. Thread-safe tracker with 18 models, real pricing data, budget alerts, and per-session breakdowns. 17 tests, CI green.

Python · LLM APIs · Cost Optimization
LLM Shelter

LLM Shelter

Input/output safety validation for LLM applications. 5 validators covering PII detection, prompt injection, toxicity filtering, and content policy enforcement. Zero-dependency core. 32 tests.

Python · Security · NLP
PromptDiff

PromptDiff

Git-style prompt versioning with semantic diff. Track prompt evolution, compare versions structurally, and catch regressions before they hit production. 22 tests.

Python · Prompt Engineering · Versioning
Bench My LLM

Bench My LLM

Lightweight LLM benchmarking toolkit. 4 prompt suites measuring time-to-first-token, tokens-per-second, and quality across models. Compare providers in minutes. 12 tests.

Python · Benchmarking · LLM APIs
Agent Replay

Agent Replay

Record, replay, and debug LLM agent executions. Rich terminal viewer, HTML export, 7 modules for full observability into multi-step agent workflows. 21 tests.

Python · Agents · Observability
MCP Forge

MCP Forge

Scaffold production-ready MCP servers in seconds. 7 Jinja2 templates generating runnable Model Context Protocol servers with tools, resources, and tests. 23 tests.

Python · MCP · Code Generation
06
06

Let's talk.

GitHub LinkedIn X