Prashant Sinha
Lead Developer
Info
- Paris, France
- 🇬🇧 English - Bilingual
- 🇫🇷 French - C1 (Advanced)
- 🇮🇳 Hindi - Bilingual
Summary
I am a software developer based in Paris. I currently lead Partoo’s AI infrastructure, where my team builds and ships agentic systems handling hundreds of thousands of requests per week. My work focuses on using programming, mathematics, and design as tools to solve interesting and challenging problems across a wide array of domains: Agentic AI, Machine Learning, Computer Vision, Scientific Computing, Web Applications, and Embedded systems. I have published peer-reviewed research in computational biology and genetics, and I maintain several open-source projects in my spare time. I also enjoy designing and printing functional 3D models.
Agentic Architecture & AI Chatbots
Leads a team of three owning Partoo’s AI infrastructure. Spec’d and developed Jim, an AI chatbot (OpenAI Chat Completions API) with a custom DSL for tool calling, memory, and context compaction. Authored the company-wide Agentic Architecture spec — framework-agnostic infrastructure for model management, usage tracking, a real-time dashboard, and conversation memory. Built on Pydantic AI, it enabled ~10 agents in production handling 300k requests/week.
Currently being reworked into an AI Gateway. Managed sprints, product specs with PM, and team repositories.
Feedback Management Analytics
Spec’d and implemented an analytics dashboard for surveys replacing a third-party tool, powered by ClickHouse. Fixed critical SQL bugs — one fix reduced query memory usage from >4 GB to <300 MB.
Business Messaging
Developed real-time integrations for Google Maps, Instagram, Messenger, and WhatsApp. Grew from executing sprint tasks to writing technical strategies, with a track record of on-time releases
WeLearn - A tool for collective learning
I designed and built the APIs that power a machine-learning model capable of classifying web pages into Wikipedia concepts, allowing users to map and explore their learning pathways. I also created an interactive concept‑map visualization that uses spatial cues to display a user’s knowledge landscape and highlight similar concepts for exploration. Finally, I developed browser extensions that capture web pages as users browse, feeding the data into the classification pipeline in real time.
Hydrodynamics of Super-Paramagnetic Micro-Beads
My work here was to capture the motion of the micro beads undergoing attraction due to an external magnetic field. Based on this it is possible to quantify hydrodynamic drag forces on the beads.
Using open source tools
trackpyandscikit-image, I developed an image processing pipeline to isolate, track, and calculate motion dynamics from the frames. We reached an average frame rate of ~800 fps.The Mechanical Sense of Paramecium
As a research intern, I developed high-frequency, real-time image analysis and hardware control pipeline to track long trajectories of Paramecium. I wrote a driver for the microscope stage, and a high-speed image aquisition and processing (~50fps) tool. Using these software we were able to follow the trajectory of moving Paramecium.
Study of Diversity in Phyllotactic Patterns
Phyllotaxis is the arrangement of lateral organs in plants. In this project we attempt to explain the origin of emergent patterns, transitions, and irregularities in such arrangements. Using cylindrical projection unrolling and unwarping, I developed an image analysis pipeline allowing better study of complicated plant organs.
Personalised Medical Events, Conferences, and Courses Recommendation for Doctors at NHS, UK
- Developed the algorithm for classification of events into medical specialities.
- Integrated the classification system into the web app.
- Implemented an event recommendation engine leveraging Elasticsearch and stored features.
- Wrote a framework for event data aggregation from multiple NHS websites.
Computer Vision and Information Extraction
Developed and shipped a utility for digitizing waveforms and extracting metadata from
ECGreports. It is being used at Brigham and Women’s Hospital, Boston for processing more than a million patient records.Also worked on extracting normalized tabular data from scanned medical records. Key challenges here involved handling skewed, desaturated and otherwise low quality images.
Survaider Web App and Survey Builder
Developed a web app backend and REST API in Python.
Designed and developed a web based survey builder tool which was later integrated into the core application.
Human Activity Recognition from Acceleration data Under Naturalistic Conditions
The goal of this project was to develop a system for recognition of low level human activities, such as Walking, Running, Sitting, etc., using a single accelerometer.
Using supervised learning method, we trained a Random Forest Classifier using feature vectors obtained via our model. We were able to produce an accuracy of
95%through our model.Modelling and Network Analysis of Intronic miRNA Mediated Gene Expression Regulation
In this project, we developed a mathematical model for gene expression regulation via intronic micro
RNAs, based on emperical data and experimental observations at IGIB.As a part of study, we’ve also generated a
miRNA <=> mRNAtarget network using various available databases.
Experience
Protein engineering and de novo designing of a biocatalyst
M Kaushik, P Sinha, P Jaiswal, S Mahendru, K Roy, S Kukreti
Journal of Molecular Recognition, 2016
Intronic miRNA mediated gene expression regulation controls protein crowding inside the cell
P Sinha, P Jaiswal, AK Jainarayanan, SK Brahmachari
Gene, 2018
Meta-analysis of genomic variants and gene expression data in schizophrenia suggests the potential need for adjunctive therapeutic interventions for neuropsychiatric disorders
SA Chellappa, AK Pathak, P Sinha, AK Jainarayanan, S Jain, SK Brahmachari
Journal of genetics, 2019
Investigating the “wisdom of crowds” at scale
A Mysore et al.
ACM UIST, 2015
Open Source
Pen Plotter with automatic Toolchanger
The plotter can use up to 5 pen/tools per session and has fully automated tool calibration. Build details available on Hackaday.
DisInfo: (Info)rmation (Dis)play
A weekend project for past couple of years, Disinfo tries to make it easier to build animated, interactive, and physical dashboards. I use it as a sandbox to explore and better understand existing technologies and patterns, such as Flex layout, declarative and composable view components, and interfacing with external APIs while maintaing performance on resource constrained devices. In my tests the system easily achieves 60fps on a Raspberry Pi 5.
PyBloomFilter: Fast, Memory Mapped Bloom Filter
The goal of
PyBloomFilteris to provide a fast, simple, scalable and correct library for Bloom Filters in Python. I’ve been maintaining this package since 2016.Hues: Colored terminal text made easy for Python
Hues is a Python package which facilitates printing ANSI encoded text to terminal in unix-like systems.
metaRNA: Find target sites for the miRNAs in genomic sequences
metaRNA finds potential target sites for the microRNAs in genomic sequences. It is built on miRanda, an algorithm for detection and ranking of the targets of microRNA.
Education
Approches Interdisciplinaires du Vivant
- Major
- Applied research expertise through placements in scientific labs
- Minors
- Research Ethics and its associated methodologies
B.Tech. (IT and Mathematical Innovations)
- Major
- Information Technology
- Minors
- Mathematics, Systems Biology
Skills
Development
- PythonExpert
- Agentic AIAdvanced
- JavaScriptExpert
- LinuxProficient
- Embedded SystemsProficient
- Distributed ComputingProficient
- Python
Design
- Systems DesignProficient
- 3D PrintingExpert
- Hardware PrototypingProficient
- UI/UXAdvanced
- Systems Design
Misc
- Scientific ComputingAdvanced
- Reverse EngineeringIntermediate
- Clean-Room DesignIntermediate
- DevOpsIntermediate
- Scientific Computing