Mridul Sarkar — Applied AI Engineer · Forward Deployed Engineer

Applications & Projects

ApplicationJun 2026 — Present

PaveAI — AI Job-Management CRM

Pave AI Inc. · getpave.ai

Job-management software for paving contractors. It tracks deals, estimates, invoices, purchase orders, and RFQs and keeps them continuously synced with QuickBooks so margin is visible mid-job. Its AI assistant walks a lead through its whole lifecycle — qualification, estimating, proposal, approval, and task execution — reading and updating CRM data and syncing QuickBooks through tools, with a human approving every change.

Architecture

React 19
Node + Express
Python + FastAPI
LangGraph
PostgreSQL
QuickBooks API

getpave.ai

ToolJun 2026

screenshot-ai

Personal project

A zero-dependency macOS bash helper that, on every screenshot, asks “Using this for AI?” and offers an auto-delete that self-destructs the file in 5 minutes — no uploads.

bash
macOS
privacy
tooling

github.com/mertall/screenshot-ai

AgentDec 2025

Hardware Agent

Personal project

An agentic system that auto-detects a machine's CPU/GPU/RAM/storage/OS and recommends the optimal local-LLM deployment for it — built on LangChain tool-calling, cross-platform across macOS and Windows. Powers the Local AI Station installer's right-sizing.

LLM
LangChain
agents
hardware
Python

ApplicationOct 2025 — Present

Enroads ERP

Enroads Paving (client deployment)

An ERP for a paving company covering CRM, project management, and financial operations — with documents and email in Microsoft 365 and continuous QuickBooks sync, so the office runs sales, projects, and books from one system.

Architecture

React
Express
GraphQL
Flask
PostgreSQL
QuickBooks API
Microsoft 365

ResearchMay 2025

CLIP Finetune on CIFAR-10

Personal project

A multimodal RAG and contrastive-learning experiment fine-tuning CLIP on augmented CIFAR-10 image/text pairs, then extracting embeddings and indexing them in HNSW for retrieval evaluation.

CLIP
multimodal
contrastive learning
HNSW
RAG

github.com/mertall/clip-finetune-cifar10

ResearchMay 2025

PatchTST + HNSW — Time-Series Vector Search

Time-series retrieval

Extended vector search to time series: trained a lightweight PatchTST on Australian electricity demand using fixed-length, leak-free windows, indexed the train embeddings in HNSW, and benchmarked brute-force KNN vs HNSW using held-out windows as queries.

Architecture

PatchTST
HNSW
Time series
Embeddings
PyTorch

github.com/mertall/patchtst_and_hnsw

ApplicationApr 2025 — Present

Local AI Station

Multimodal Works

A private AI workstation that runs entirely on a client's own hardware — no data leaves the building. ~20 open-source services compose via Docker Compose profiles: a Hermes coordinator routes to agent tooling (OpenWork/OpenCode, Claude Code + Codex) over a LiteLLM gateway whose classifier router fronts Ollama-served local models (Qwen3 / Gemma), with opt-in metered cloud routes. Memory spans three tiers under one entity identity — Postgres relational, Apache AGE knowledge graph, and Qdrant vectors — fed from an Obsidian knowledge vault. Caddy reverse-proxies the UIs, n8n drives automation (GitHub PR webhooks via smee.io), and Langfuse traces everything.

Architecture

Docker Compose
Ollama
LiteLLM
Postgres + Apache AGE
Qdrant
Obsidian
n8n
Caddy
Langfuse

ExperimentApr 2025

MPS Diffusion Model

Personal project

An exploratory diffusion-model implementation with Apple-Silicon MPS acceleration in PyTorch.

diffusion
PyTorch
MPS
generative

HackathonApr 2025

Quantum Computing Hackathon — 8th of 600 teams

BlueQubit

Optimized a 60-qubit tensor-network contraction under hardware constraints, placing 8th of 600 teams.

Quantum
Tensor networks
Optimization

github.com/mertall/bluequbit_hackathon

ProjectMar 2025

FellowshipFinder

Personal project

An AI agent that lets graduate students upload a resume and surfaces matching fellowships at UChicago, with a frontend/backend split running OpenAI or a local Mistral-7B model.

LLM
agents
RAG
OpenAI
Mistral

github.com/mertall/FellowshipFinder

Project2025

Water Bottle Challenge — Spectrogram K-means

Audio Classification

Classifies top-vs-bottom strikes on a water bottle from their audio spectrograms — an unsupervised K-means classifier reaching 95% blind-test accuracy through robust preprocessing and careful evaluation.

scikit-learn
Audio
K-means
Signal processing

github.com/mertall/audio-classification

ProjectNov 2024

Multimodal Image–Text Vector Search

VisionSearch · CLIP · HNSW · FastAPI · SageMaker

Text-to-image search over CLIP embeddings with an HNSW (hnswlib) index, served via FastAPI. A CLIP model is packaged with a custom handler, pushed to S3, and deployed as a SageMaker inference endpoint returning L2-normalized embeddings. Chose embedding retrieval over generative RAG for latency and determinism.

Architecture

CLIP
hnswlib
FastAPI
SageMaker
S3
PyTorch

github.com/mertall/VisionSearch

Deployed Solution2021 — 2025

Enterprise Data Platform

Accenture Federal Services

A production system serving 100k+ users at 99.9% SLOs with p95 latency cut 40%. I built the full backend — a streaming API, data ingestion, processing, and serving APIs — with authentication handled by the DoD client. I built the Elasticsearch indexing pipelines with NLP enrichment that boosts discoverability: back- and forward-translation, synonym augmentation, and generative text creation. I also architected and built a multi-tenant, cross-organization user-profile service that tracks preferences and metadata from external orgs, and built an MCP server exposing the platform's APIs behind an AI agentic chat interface to explore and understand new data and to upload new data. Architecture shown high-level (proprietary/federal).

Architecture

Data platform
Streaming API
Elasticsearch
NLP enrichment
MCP server
Agentic chat
Multi-tenant
SLOs
Deployed

Prototype2022 — 2024

GenAI Retrieval Prototype (RAG)

Accenture Federal Services

An early production bet on vector search, adopted as the technology first emerged: documents embedded with Sentence Transformers into an Elasticsearch vector DB, retrieved for RAG, improving Recall@10 ~20%. Informed an LLM-deployment white paper for the Intelligence Community. Architecture shown high-level (proprietary/federal).

Architecture

RAG
Sentence Transformers
Elasticsearch vector DB
Evals
Early adopter

Deployed Solution2021 — 2025

Federated Search Platform

Accenture Federal Services

A production federated search with ACID-compliant NoSQL queries (MongoDB multi-document transactions) that federates across on-prem and cloud databases — JDBC ingest into MongoDB, NLP enrichment, Elasticsearch indexing — served by distributed microservices to 2M+ users. Architecture shown high-level (proprietary/federal).

Architecture

ACID NoSQL
MongoDB
Elasticsearch
NLP
Microservices
On-prem + Cloud
Deployed

ProjectMay 2020

HOSVD Facial Recognition

Personal project

A higher-order SVD facial-recognition project intended to drive room security and mood lighting, combining OpenCV/PiCam capture with IR-sensor scripts on a Raspberry Pi.

Computer vision
HOSVD
OpenCV
Raspberry Pi

github.com/mertall/HOSVD

ProjectMay 2019

QEOT — Qualitative Evaluation of Text

Personal project

An early NLP project that cleans web-page text into a knowledge base — two preprocessing models plus a custom “understandability” metric, built with spaCy and CoreNLP while researching SQuAD 2.0 and BERT for downstream QA.

NLP
spaCy
CoreNLP
BERT
Python

github.com/mertall/QEOT