Hi! I'm a software developer who works on backend systems and the infrastructure behind AI agents. I like exploring and tinkering with new things, and meeting new people along the way. Outside of code, I love learning about astronomy.
I'm open to opportunities in backend engineering and applied AI.
Experience
Associate Backend Developer - Agentic AI Integration
January 2025 — PresentCommotion(acquired by Tata Communications)/Full-time/Mumbai, India
- Engineered the integration and tool-execution layer that Commotion's Voice AI agents call to act inside customers' real business systems. Built and shipped 100+ enterprise connectors exposing 3,200+ agent-callable actions and 325 webhook event triggers across 17 applications, spanning SAP, Oracle Fusion, Salesforce, ServiceNow, Workday and 95+ other enterprise platforms.
- Delivered integrations end to end for 10+ named enterprise accounts across Tata group companies and multinationals, from credential negotiation with the customer's IT team through to production release: a contact-centre Voice AI bot and an Accounts Payable invoice automation pipeline for Tata Chemicals over SAP S/4HANA and Ariba, IVR-facing actions for Genpact, and 22 customer-written Python scripts ported into standard platform actions for HDFC ERGO.
- Built MCP and A2A agent infrastructure on mcp-context-forge, the Python gateway fronting every tool call an agent makes. Extended external A2A agents to full OAuth by carrying credential references through the gRPC interface instead of raw secrets, resolved at run time, and wrapped an external MCP server as a first-class platform connector with credential verification that fails at creation rather than at call time.
- Load-tested the MCP gateway under simulated voice-call concurrency with a harness that holds persistent SSE sessions through the MCP protocol handshake and issues tool calls on a conversational cadence. Swept 40 to 500 concurrent sessions against pod count, Gunicorn workers, CPU limits and tool type, capturing p50 to p99 tail latency to set production pod sizing and the threshold for scaling horizontally rather than vertically.
- Diagnosed and fixed an OAuth flow that broke under pod autoscaling. The SSE session was held in local process memory, so the provider's callback could land on any pod while only one held the client's connection, a bug invisible at one replica and guaranteed at scale. Replaced the in-memory store with Redis Pub/Sub keyed by session ID to decouple session ownership from request routing.
- Optimised the gateway's hottest query paths by adding composite (name, enabled) indexes to the tool and A2A agent tables through Alembic, the predicate executed on every single tool call and agent invocation, while deliberately excluding a table whose lookups are always by primary key. Replaced per-request Postgres clients with cached connection pools that evict themselves on fatal errors.
- Engineered enterprise authentication across OAuth 2.0 authorization-code and client-credentials flows, including in-app SHA-256 SAML bearer assertion signing for SAP SuccessFactors, which issues no refresh token, a multi-tenant fix replacing a hardcoded Atlassian cloudId that routed every customer's requests to a single tenant, and credential-scoped action maps so an agent only sees the actions its credential can run.
- Hardened the platform's security and supply chain: cleared Trivy and Black Duck Polaris findings through non-root container execution, read-only root filesystems, dropped capabilities and dependency patching; eliminated a raw interpolated-SQL write path in favour of parameterised queries over a pooled connection; and drove lockfile pinned builds so dependency versions cannot drift between environments.
- Promoted from Backend Developer Intern (January 2025 – June 2025).
Full Stack Developer Intern
June 2024 — July 2024Arxena Inc./Internship/Mumbai, India
- Built a GPT-powered WhatsApp recruitment chatbot in NestJS, GraphQL and WebSocket API that talked to candidates on the recruiter's behalf. It asked preliminary screening questions about the job description and collected the answers, sent the JD as a PDF whenever a candidate asked for it, and collected resumes over WhatsApp into Amazon S3.
- Handled voice replies as well as text: when a candidate answered with an audio message, the bot fetched the audio file, transcribed it and stored the transcribed response.
- Built a WhatsApp-style chat interface in React.js inside Twenty CRM so the recruiting team could follow each candidate's conversation, with the candidate's status from the CRM shown alongside the chat.
- Cut the CRM server's AWS EC2 cost by 75%. The NestJS (Twenty CRM) app was served and built on the same t2.large, so every CI/CD build slowed the live app and the instance sat oversized between builds. Moved serving to a t2.small and wrote a bash deployment script that launches a t2.large from an AMI only for each build, copies the compiled output to the server over SCP, then terminates the build instance.
- Consolidated the Python backend from four smaller EC2 instances into a single t2.large, a further 10.8% saving.
Information Technology Intern
June 2023 — September 2023Kraymera Study Abroad/Internship/Mumbai, India
- Implemented CRM systems and processes using Zoho and built WhatsApp Business API automation in Deluge, boosting lead nurturing and conversion rate.
- Installed and configured CentOS on a server and set up networking devices and a GSM gateway for telecommunications.
Education
Bachelor of Engineering in Information Technology
2021 — 2025Thadomal Shahani Engineering College (TSEC), University of Mumbai/Mumbai, India/CGPA 8.54 / 10
Open Source
Linux Kernel Mentorship Program Mentee
March 2026 — August 2026The Linux Foundation — LFX/Spring 2026
- Graduated with 11 patches contributed to the Linux kernel mainline, and published a blog sharing the experience.
- Converted text-based Devicetree bindings to JSON Schema, and fixed documentation links, typos and driver error handling.
- Worked within the kernel development workflow, patch submission and review cycles, using cscope, Coccinelle and smatch for static analysis.
- Selected for the programme from 693 applicants.
11 patches merged into mainline: Devicetree bindings converted to JSON Schema, plus documentation and driver error-handling fixes.
Contributions to the workflow automation platform.
Contributions to the open-source digital logic circuit simulator.
Projects
Research
Evaluating Human Speech Confidence with Deep Learning Techniques
September 2025Ninad Naik et al./ICCIS 2025 - BITS Pilani, Goa
An automated system that classifies perceived speaker confidence on a five-point scale from audio alone, built to make confidence assessment and feedback scalable.
- Framed the task as five-way classification of perceived confidence from speech, where the hardest problem is that there is no ground truth to measure against: the dataset is the definition of the target, so every choice about speakers, labels and clip length changes what the model learns to detect.
- Seeded a dataset of 751 thirty-second audio clips with no recording budget, drawing from YouTube, TED and TEDx talks and recordings by seven volunteers, and choosing sources where the confidence label is inferable from context: early videos of creators who later became well known for the low-confidence class, rehearsed conference speakers for the high-confidence class. The result spans a far wider range of ages, genders, accents and recording conditions than a self-recorded set could.
- Had four annotators rate every clip on a 1 to 5 Likert scale and took the mode as the final label, measuring agreement at a mean Cohen's Kappa of 0.6203 and a Fleiss' Kappa of 0.6193. That figure is the practical ceiling for the task: four humans labelling the same clip agree only to that degree, so this is genuinely subjective rather than simply noisy.
- Compared traditional acoustic features against learned ones, extracting MFCCs alongside Wav2Vec2 embeddings from wav2vec2-base-960h, taking the (T, 768) hidden states through temporal mean pooling into a single 768-dimensional vector and standardising with StandardScaler.
- Evaluated Logistic Regression, a multilayer perceptron and a Fuzzy Hypersphere Neural Network under 5-fold stratified cross-validation, reducing the Wav2Vec2 features from 768 to 128 dimensions with SelectKBest on the ANOVA F-value for the FHNN.
- Wav2Vec2 with an MLP gave the best exact accuracy at 0.59, FHNN with MFCCs the best macro F1 at 0.62, and both beat the Logistic Regression baseline at 0.47. The non-linear models reached 0.94 to 0.96 relaxed accuracy, counting a prediction within one confidence level as correct, so predictions land close to the human label even when not exact.
Skills
Languages
- TypeScript
- JavaScript
- Python
- Go
- C
- C++
- Java
Backend & APIs
- NestJS
- FastAPI
- Node.js
- GraphQL
- gRPC
- OAuth2.0
- JWT
AI & Agents
- MCP
- A2A
- LangChain
- LangGraph
Data & Messaging
- PostgreSQL
- MongoDB
- Redis
- MySQL
- Kafka
Infrastructure
- Linux
- Docker
- Kubernetes
- AWS
Frontend
- React.js
- Tailwind CSS
- HTML
- CSS