Now · Technical Product Manager, Agent Fleet at Acuity Analytics

Parag Chaudhari

Technical Product Manager and AI builder

5+ years across machine learning, ML platforms and agentic AI. I lead the product and its cross-functional team on Agent Fleet, an agentic AI platform for enterprise financial-services workflows, and I also ship features myself. Before that I built and shipped RAG, LLM-evaluation and demand-forecasting products at RapidCanvas and aioneers, and I build open-source AI and developer tooling end to end in my own time.

linkedin.com/in/pcihd github.com/pcihd
Selected results
98%

Accuracy on a Gurumukhi RAG solution that preserves spiritual texts.

RapidCanvas
80%

Fewer operational errors from a logistics process-optimization solution.

RapidCanvas
40,000+

SKUs covered in one FMCG demand-forecasting delivery, at a 6% value add.

aioneers
4 mo to 3 wk

Implementation time for an automotive forecasting engine.

aioneers
20%

Value add over the customer forecast at the automotive manufacturer.

aioneers
Skills

Product

RoadmappingJiraRelease managementStakeholder management

Agentic AI

Agent workflowsRAGLLM evalsClaudeOpenAI

Local AI

vLLMllama.cppLiteLLMWan2GP / LTX

ML

Time-series forecastingLightGBMPyTorchComputer vision

MLOps

OptunaMLflowLangfuse

Engineering

PythonTypeScriptFastAPIReactTextual

Cloud

AzureDatabricksAWS
Experience

Acuity Analytics

May 2026 - Present

Technical Product Manager, Agent Fleet

  • Own product direction and delivery of Agent Fleet, an agentic AI platform for enterprise financial-services workflows, from requirements through release.
  • Hands-on, not just managing: I ship features directly alongside the engineering team, in addition to running the backlog, priorities and releases.
  • Coordinate engineering, QA, data science and UX on a single Jira board, aligning priorities, dependencies and release readiness across teams.
  • Manage stakeholder expectations, risks and escalations, turning platform progress into clear status for leadership.

RapidCanvas

Jan 2025 - May 2026

Product Data Scientist

Platform development
  • Brought a data science perspective to product development and contributed to scoping and specifying MLOps feature requirements.
  • Built an in-house RAG package that reduced time to MVP to a day.
  • Contributed UI/UX changes that cut the experimentation iteration cycle for RAG use cases by 20%.
  • Managed product releases and worked cross-pod to ship on time.
  • Built a curated set of “Sample solutions” with prebuilt components for common problems, saving up to a week in every implementation.
Enterprise solution delivery
  • Delivered a computer-vision solution for transportation corridor maintenance for an EMEA customer.
  • Built a process-optimization solution for a logistics company and GCCs, increasing leadership visibility and reducing operational errors by 80%.
  • Solved invoice-to-PO matching for affiliate media networks and GCCs using fuzzy matching and LLM-based document extraction.
  • Delivered a RAG solution in Gurumukhi for preserving spiritual texts, at 98% accuracy.
  • Built a robust LLM evaluation approach for domain-specific answers, combining LLM-as-judge prompting with human evaluators to improve accuracy and relevance.

aioneers

Sept 2021 - Jan 2025

Senior Data Scientist

Demand forecasting: client projects
  • FMCG manufacturer: delivered demand forecasting across 40,000+ SKUs, with a 6% value add over the customer forecast.
  • Automotive manufacturer: implemented a forecasting engine with a 20% value add, cutting implementation time from 4 months to 3 weeks.
  • Created features for product life-cycle management through machine learning.
Demand forecasting: product
  • Built a demand forecasting engine as a decision-support product and led a team of data scientists across client projects and product development.
  • Proposed and implemented a model-agnostic architecture supporting a wide bouquet of models, with Optuna for tuning and MLflow for experiment tracking.
  • Built feature factories to consume exogenous variables such as holiday calendars, purchase orders and customer forecasts.
  • Built an end-to-end forecasting-as-a-service pipeline, from low-cost data ingestion to forecast quality recommendations, with dashboards for operational and strategic personas.
  • Mapped planners to forecasts and used the recommendations to reduce planner workload.
Copilot for supply chains and internal library
  • Built an LLM copilot for supply-chain use cases, including hosting for high workloads, abstractions over LLMs and indexes, and algorithms to consume large indexes within model token limits.
  • Proposed and developed an agent framework with intelligent decision-making for complex business problems, with RAG over index databases.
  • Created an internal Python library encapsulating business logic such as dead-stock analysis, reducing custom scripting and improving reliability across projects.

Earlier experience

2019 - 2021
  • DRDO, AI Intern (Jul 2019 - Feb 2020): robotic pick-and-place with ROS, a UR10 arm and Kinect 3D point clouds, using PyTorch to identify objects and plan grasps.
  • Freelance ML (Dec 2020 - Jan 2021): a CNN with skip connections reaching 85%+ on CIFAR-100, an LSTM for particulate-matter prediction deployed in an Android app, and a CNN lung-scan classifier.
Projects

VideoForge self-hosted AI video production

Jul 2026 - Present

One idea, topic or ebook in, a finished, narrated, scored video out, on a single RTX 5080 with no cloud APIs. At the centre is Leonardo, an agent that runs the whole production from one chat message. Built end to end with Claude Code.

Painting the scenesprompts and first frames
Researching the factsweb dossier, real people
Rendering with the caststills, then video clips
Video done!with its YouTube pack

Leo is Leonardo's mascot on the chat page: pure CSS and SVG, zero GPU cost. He changes hats by project mode and reacts to each stage that lands or fails.

Stills and clips are rendered with Wan2GP. The YouTube pack is the title, description, tags and thumbnail. Tool names are Leonardo's real ones.

Featured agent

Leonardo: one prompt in, a finished video out

A chat agent built on multiple agent harnesses. It used to take a dozen screens to make a video; now you type an idea and Leonardo walks the whole journey, either guided, where it stops and asks when a choice is yours, or fully autonomous. Each workspace is one YouTube channel or series, with its own defaults, cast and tool switches.

Thinks by stageDeep reasoning for the story arc, medium for prompts, none for the YouTube copy. Slow thinking only where it improves the result.
Researches real topicsBuilds a dossier from web search, finds the real people involved and casts them from pinned photos.
Guided or autonomousSwitch mid-run, or send a note while it works.
Survives restartsAn interrupted run resumes from its transcript, and a finished one takes follow-ups.
Guardrails live in the tools, not the prompt

LLMs are good at deciding what to do next and bad at knowing what they are about to destroy. Each tool is the same handler a UI button calls, so every guard applies to the agent. I watched real runs and closed the holes one by one:

  • Wasted work. A model re-ran the story generator just to check the run time and threw away 26 minutes of written prompts.Rebuilds are refused unless asked for, and the refusal spells out what would be lost.
  • Finishing early. A model read "project created" as "video underway" and tried to finish.It cannot finish until a final video exists, and every status check names the next step.
  • Invented research. With web search down, a run carried on from the model's guesses and cast an AI model as the human designer.No search, no research: the run stops and says how to fix it.
✦ Leonardo · guided runexample
Make a 3-minute explainer on how tides work. Calm tone, six scenes.
researchthinking: medium12 sources read, 5 facts saved
build_storythinking: deep6 scenes, about 3:04 of narration
build_promptsthinking: medium6 scene prompts written

Scene 1 opening frame is ready. Keep it?

KeepRender againSkip
Keep. Then check the total run time.
build_storyRefused: 6 written prompts would be discarded. Pass rebuild=true to confirm.
Understood, I will read the run time from status instead. Next: render.

Illustrative example with an invented conversation, not a real session.

Size of the codebase, to scale
Backend: Python, FastAPI, SQLite65K
Frontend: React, TypeScript, Vite24K
Tests: 191 files, backend and frontend73K
578commits since Jul 2026
207merged PRs
30+single-purpose agent modules
6project modes
7switchable local LLMs
1GPU runs it all
  • Agent system. About 30 single-purpose LLM agents (seeds, cast, story, prompts, a supervisor that improves prompts, continuity, music, packaging, vision review) with per-mode and per-model system prompts, across six project modes including multi-character, research and ebook-to-video.
  • One GPU. Seven switchable local LLMs behind LiteLLM (vLLM, llama.cpp, a 125B mixture-of-experts model with its experts in host RAM), a VRAM and RAM handoff with GPU leases designed from real OOM failures, and per-call telemetry with a cost view that prices the local workload at OpenRouter rates.
  • Reliability. The Wan2GP driver uses named-parameter job bundles after an upgrade silently broke numbered fields, progress is derived from saved data, and tests run against a seeded backend and a fake renderer, so no GPU is needed.
  • Process. Built with Claude Code, with ARCHITECTURE and HANDOVER docs kept current alongside the code.

sys-tui terminal hardware visualizer

Python, Textual · 2026
● live◐ derived○ spec ▲ toward CPU▼ away

Simulated sensors, like the app's --demo mode. Click a part, or pick one below.

  • Built systui and gputui, two terminal apps that draw a machine's motherboard and its RTX 50-series GPU PCB as true-colour pixel art, animated from live sensors, plus pixboard, a shared canvas, chart and PNG library. About 21k lines of Python across 111 files, Apache-2.0.
  • Tagged every value as live, derived or spec so users can tell measured data from inferred, with click-to-inspect on any component and 12 RTX 50 board profiles.
  • Built an IPC map of processes, containers, VMs, LAN peers and Internet endpoints on a pure-Python netlink collector, with no extra dependencies and throttled polling. Added GDDR and PCIe bandwidth overlays on the GPU's PCB traces.
  • Designed phase 2, a read-only systui-to-systui agent for viewing VMs and LAN machines from one instance.

CPU

    Education

    MIT ADT University

    Bachelor of Science in Computer Science and Engineering, Specialization in Artificial Intelligence

    Earlier projects · 2021 - 2026
    • Vibe-coded a content-generation tool (LLM scripting, curation and validation, Gemini Veo 3 visuals) and a Lofi MIDI music mixer with prompt-based music generation (Lyria RealTime).
    • Dreambot: a GAN that maps solid-fill colour regions to realistic image components, built in PyTorch.

    Web copy. Phone and email are left off this page; message me on LinkedIn. Glass styling inspired by the ASMR view in VideoForge.