Devansh Verma

I build machine-learning systems and run them on real infrastructure. The thread through my work is restraint: a rules layer instead of a rebuilt scoring model, a human approving every automated dispatch, fairness reported as a full range instead of one flattering number. I would rather ship the version that holds up than the version that only demos well.

1. ▲ Call 1A rules layer instead of a rebuilt scoring modelP4 Customer-health triage engine →
2. ▲ Call 2A human approving every automated dispatchP1 Atlas — emergency-dispatch copilot →
3. ▲ Call 3Fairness reported as a full range instead of one flattering numberP2 Fairness stress-testing framework →

1. P1Hackathon · 2-day build · 2026

   Atlas — emergency-dispatch copilot

   A copilot for 911 dispatch: it transcribes incoming calls, clusters duplicate reports into single incidents, and recommends which units to send.

   ▲ Team radio · the call

   The system only recommends — a human dispatcher approves every dispatch. The optimizer re-plans the moment a higher-priority call lands, instead of treating allocation as a one-shot decision.

   45 units and 3 hospitals simulated with live capacity tracking; 18 tests; built in two days.

   45 units

   3 hospitals

   • 2-day build
   • 18 tests
   • A human approves every dispatch
   • FastAPIFastAPI A Python framework for building web APIs.
   • SQLAlchemySQLAlchemy A Python toolkit for working with databases.
   • ReactReact The most widely used library for building web interfaces.
   • LeafletLeaflet A JavaScript library for interactive maps.
   • WebSocketsWebSockets A live, two-way connection between a browser and a server.
   • Docker ComposeDocker Compose Runs several Docker containers together as one app.
   • LLMsLLMs Large language models: the kind of AI behind chat assistants.
2. P2Team lead · graduate ML · 2026

   Fairness stress-testing framework

   A framework that tests whether a published fairness fix actually holds up when you change the random seedrandom seed The starting number for a model's randomness; changing it can quietly change the results., the decision thresholddecision threshold The cut-off score above which a model says yes., or the subgroup you look at.

   ▲ Team radio · the call

   Reported the full range of results across random seeds and thresholds instead of the usual single flattering number — and showed that how fragile a fix is depends on the model.

   Threshold choice alone swung the fairness gapfairness gap A fairness measure: the difference between groups in how often a model gives the positive outcome. 0 means every group gets it at the same rate. 7× (0.04 → 0.28), as much as switching the debiasing methoddebiasing method A technique applied to a model to reduce unfair differences between groups..

   7×Fairness gap · threshold alone

   0.040.280.000.100.200.30
   • scikit-learnscikit-learn The standard Python library for classic machine learning.
   • XGBoostXGBoost A popular library for gradient-boosted decision trees, strong on tabular data.
   • FairlearnFairlearn An open-source toolkit for measuring and improving fairness in machine-learning models.
   • pandaspandas Python's main library for working with tables of data.
   • NumPyNumPy Python's core library for fast numerical arrays.
3. P3Applied cryptography · 2026

   Secure login: TLS 1.3 + OPAQUE, from scratch

   A TLS 1.3TLS 1.3 The protocol behind the padlock in your browser: it sets up an encrypted connection to a website. 1.3 is the current version.–style secure connection built from scratch in Python, with OPAQUEOPAQUE A modern password-login protocol: the server can check your password without ever seeing it or storing it. password login in place of the usual Diffie-HellmanDiffie-Hellman The classic way for two computers to agree on a secret key over an open network. key exchange: the server never sees, stores or can pre-compute against your password.

   ▲ Team radio · the call

   Ran the OPRFOPRF Oblivious pseudorandom function: the server helps turn the password into a key without learning the password. on plain Ed25519Ed25519 A widely used elliptic curve, behind many digital signatures and SSH keys. curve operations instead of the textbook Ristretto255Ristretto255 A safer way of using the same curve that removes a known pitfall (the cofactor); the textbook choice for OPAQUE. — safe here because every OPRF output is hashed, so no raw curve point is ever exposed.

   A wrong password fails on your own device and leaks nothing to the server. 12/12 tests pass.

   ClientPassword

   ✕

   Server

   • never sees it
   • never stores it
   • never pre-computes against it

   12/12 tests pass
   • PythonPython The default language for machine learning, data and back-end work.
   • Ed25519Ed25519 A widely used elliptic curve, behind many digital signatures and SSH keys.
   • X25519X25519 Diffie-Hellman key agreement on Curve25519, the standard in modern TLS.
   • AES-GCMAES-GCM AES in a mode that both encrypts data and detects any tampering with it.
   • HKDFHKDF A standard way to turn one shared secret into several separate keys.
   • X.509X.509 The certificate format websites use to prove who they are.
4. P4Software Engineer Intern · 2026

   Customer-health triage engine

   A triage layer on top of an existing customer-health score that tells success managers which accounts need attention, why, and what to do next.

   ▲ Team radio · the call

   Built it as deterministic rules, not ML — the logic didn't need a model, and a model would have been harder to trust and to explain to the people acting on it.

   Surfaced quietly at-risk accounts the existing score never flagged.

   Customer-health score

   Deterministic rules

   Which accounts

   Why

   What next

   Rules, not ML: easier to trust and explain

   • PythonPython The default language for machine learning, data and back-end work.
   • pandaspandas Python's main library for working with tables of data.

1. 2026 – present

   Student Engineer, Engineering Technology · ASU Enterprise Technology

   DevOpsDevOps Automating how software is built, tested, shipped and run. and infrastructure work in the university's central technology organization, inside its GitHub org and AWS environment.

2. 2026

   Software Engineer Intern · ParentSquare

   Built the customer-health triage engine above, plus CRUDCRUD Create, read, update, delete: the basic operations behind most app features. feature work on the Digital Business Review product in React and TypeScript on a pull-request CI/CDCI/CD Automated testing and deployment on every code change. workflow.

3. 2025

   AI Engineer · KLJ Group

   Built and ran LLMLLM Large language models: the kind of AI behind chat assistants. inferenceinference Running a trained model to get its answers, as opposed to training it. pipelines for three products — billing automation, an email-thread reader, and voice + vision search — with Python and Go on the serving path and local and cloud models served through vLLMvLLM An open-source engine for serving large language models quickly and cheaply..

4. 2024

   AI / ML Engineer · Times Internet

   Shipped ML inferenceinference Running a trained model to get its answers, as opposed to training it. APIs and content-classification pipelines; KafkaKafka A system for streaming events between services reliably. for async events, RedisRedis A very fast in-memory data store, often used as a cache. for caching, Docker and CI/CDCI/CD Automated testing and deployment on every code change., deployed across AWS and GCP.

5. 2023 – 2024

   Hardware Benchmarking, Central Testing Team · 91mobiles

   Benchmarked and analyzed phones across CPU, GPU, memory and thermals, and trained new interns on the process.

6. 2023 – 2024

   Founder · JobCadet

   AI recruitment startup; raised ₹6 lakh in NASSCOMNASSCOM India's national IT-industry association. seed funding.

• Detection of Open Metal Sites in Metal–Organic Frameworks Using Machine Learning

  ICICC 2023, Springer LNNS · 2023

  Read →

• Vision-Based Human Activity Recognition Using CNNCNN Convolutional neural network: a model type built for images and signals. and LSTMLSTM A neural network with memory, suited to sequences such as video frames or speech. Architecture

  Advanced Computing (IACC 2023), Springer CCIS · 2024

  Read →

• Multi-featured Speech Emotion Recognition Using Extended Convolutional Neural Network

  Advanced Computing (IACC 2023), Springer CCIS · 2024

  Read →

• 2021

  NASA Space Apps Challenge

  Ranked #1 in India — led backend work turning NASA datasets into visualizations.

  2

  1

  3

  P1 · India
• 2023

  BVEST Hackathon

  Led a 100-student organizing team; 350+ participants and ₹3 lakh in sponsorship raised.

  100 organizing team

  350+ participants+

  ₹3 lakh sponsorship · one mark = ₹1 lakh₹₹₹
• 2022

  ISTE Club — General Secretary

  Led 150+ students across events and technical initiatives.

  150+ students led+

• PythonThe default language for machine learning, data and back-end work.Proven at: KLJ pipelines · fairness framework · triage engine
• GoGoogle's language for fast, simple network services.Proven at: KLJ serving path
• CA low-level language that runs close to the hardware, used here for raw speed.Proven at: AES cryptanalysis
• React · TypeScriptThe most widely used library for web interfaces, written in typed JavaScript.Proven at: ParentSquare DBR · Atlas
• FastAPIA Python framework for building web APIs.Proven at: Atlas
• vLLM / LLM servingRunning large language models in production; vLLM is an open-source engine that serves them fast.Proven at: KLJ inference products
• PyTorch · scikit-learn · XGBoostPython's core machine-learning toolkits: deep learning, classic ML and gradient-boosted trees.Proven at: research · fairness framework
• DockerPackages an app with everything it needs, so it runs the same anywhere.Proven at: Times Internet · Atlas
• Kafka · RedisKafka streams events between services; Redis is an in-memory store, often used as a cache.Proven at: Times Internet pipelines
• AWS · GCPAmazon's and Google's cloud platforms.Proven at: Times Internet deployments
• Applied cryptographyBuilding real security (encryption, key exchange, secure login) and breaking weakened versions of it.Proven at: TLS 1.3 + OPAQUE · AES CTF
• D3.jsA JavaScript library for interactive data visualisation.Proven at: interactive data-viz coursework