Sohhom Bandyopadhyay, PhD

Independent Researcher — Quantitative Trading

• Building an end-to-end research and trading system for Indian equities (NSE/BSE): tick and OHLCV data ingestion, feature and label engineering, signal generation, walk-forward backtesting, and execution logic.
• Reusable components for transaction-cost modelling, strategy evaluation, and risk metrics; every idea is put through a cost-and-survivorship break-even test before anything else.
• Emphasis on regime awareness and selectivity — deciding when not to trade is treated as part of the strategy, not an afterthought.
• Stack: Python, Polars, pandas, NumPy, SciPy, scikit-learn, broker APIs.

Applied AI Engineer (Trading Systems)

• Built LLM-powered trading copilots that interpret live chart context and answer interactive market-related queries for exploratory trader workflows.
• Designed and implemented derivatives (F&O) analytics dashboards giving quantitative and visual insight into Open Interest dynamics, Max Pain, implied volatility conditions, and options positioning.
• Developed strategy-selection systems for multi-leg options strategies, driven by market-regime indicators such as VIX conditions and contango / backwardation behaviour in the term structure.
• Worked across ML, LLMs, quantitative analysis, and algorithmic modelling to deliver prototypes and trader-assistance systems end-to-end.

Artificial Intelligence Researcher

• Reported to the CEO as a senior individual contributor; occasional collaboration with product teams.
• Research on human preference elicitation for AI alignment — how the interface used to collect preference data changes the data you get, across rating scales, pairwise and n-alternative forced choice, like/dislike and best-worst scaling.
• Built instrumentation for browser-based preference experiments, including a custom n-alternative forced-choice plugin (2-, 4- and 6-choice) that did not exist in the framework being used.
• Ran crowdsourced studies end to end: stimulus generation, deployment, participant-quality screening, and analysis in R (within-subjects factorial ANOVA with effect sizes; Bradley–Terry and Elo ranking models).
• Diagnosed a disconnected comparison graph in existing preference data that made global ranking unidentifiable, and designed the task change that fixed it.
• Synthesised preference-elicitation literature across psychometrics, experimental economics, computational social choice and RLHF.
• Studied emergent properties of LLMs (GPT-4, Claude) and their implications for EdTech; scoping and feasibility analysis including automated essay scoring.

R&D Researcher (Human–AI Interaction)

• Built and evaluated prototypes on GPT-3, ChatGPT and GPT-4 as they were released, including deployed web applications with API integration, content moderation and usage logging.
• Fine-tuned GPT-3 base models and characterised the practical limits of doing so; ran early Stable Diffusion and diffusion-model training experiments.
• Designed and deployed the first round of browser-based preference experiments on crowdsourcing platforms.
• Investigated the academic and commercial impact of these LLMs.

Doctoral Researcher — Quantifying the UX of noise control headphones

• Deep dive into signals and control systems and how they relate to human sound perception.
• Developed psychoacoustic assessment protocols for Active Noise Control headphones, including methods for measuring envelope distortion.
• Conducted human experiments measuring speech perception through ANC headphones.
• Coursework: Computational Neuroscience, Natural Language Processing, Computer Systems, Cognition and Emotion, Artificial Intelligence.
• Full thesis PDF: Psychoacoustic assessment of ANC headphones (3.9 MB).

Graduate Teaching Fellow

• Co-instructor on CG 501, a core course for graduate students in cognitive science, with Prof. Krishna Prasad Miyapuram.
• Independently designed curriculum, delivered lectures, and conducted exams.
• Taught fundamentals of machine learning and game theory.

Summer Research Intern

• Application of game theory to wireless network parameter optimisation.

Meta Quant — a machine that searches for trading strategies

• Designed a non-Turing-complete DSL for intraday strategies with validation rules that reject unsafe specs before any data is touched
• Built an automated two-layer signal-and-strategy search over NSE tick data; logged 163 experiments with full provenance
• Established that cross-sectional information coefficient does not imply tradeable within-stock edge — the search's headline metric was measuring the wrong thing
• Stack: Python, Polars, Lark, YAML, LLM coding agents

Making chart patterns measurable

• Specified informal chart-pattern vocabulary as testable geometric criteria
• Built pattern detectors under a precision-over-recall objective for human-in-the-loop use
• Evaluated motif-discovery and shape-distance methods against rule-based geometry
• Designed symbolic reduction of numeric chart state into discrete facts for LLM consumption
• Stack: Python, NumPy, SciPy, STUMPY, DTW, JavaScript

Pattern reuse, from file systems to neural networks

• N-gram frequency analysis over raw binary
• cross-file pattern overlap
• dictionary coverage growth curves
• pointer-overhead compression modelling
• sub-byte bit-level overlap
• bit-shift alignment sensitivity
• Stack: Python 3 (NumPy, mmap for zero-copy binary handling, collections.Counter, struct), JSON result serialisation

Quant investigations into Indian equities (NSE/BSE)

• Tick-level data pipelines and feature/label engineering over NSE/BSE equities
• Exact transaction-cost modelling and break-even analysis as a first-pass filter on every idea
• Walk-forward backtesting with purge gaps and symbol-day-grouped bootstrap inference
• An oracle harness that grades exit rules independently of any entry signal
• A multimodal sequence transformer for intraday cross-sectional direction
• Stack: Python, Polars, NumPy, PyTorch, LightGBM, scikit-learn

A toolbox for collecting human preference data

• Built a custom n-alternative forced-choice plugin for a web psychophysics framework, covering 2-, 4- and 6-choice image selection
• Ran crowdsourced preference studies end to end — stimulus generation, deployment, participant screening, analysis
• Implemented and compared Bradley–Terry, Elo and preference-ratio scoring on the same comparison data
• Deployed a public 4-AFC collection app with a live results dashboard

Measuring whether image generators can depict an emotion

• Ran a within-subjects study comparing emotional alignment across three text-to-image generators
• Re-derived the full three-way repeated-measures ANOVA in R with effect sizes and corrected post-hoc tests
• Correlated a published image-preference reward model against human ratings on the same stimuli
• Archived ~14.5k generations from a family of language models shortly before deprecation

Characterising ANC headphones through the playback signal

• Literature synthesis across four disciplines
• problem scoping
• psychoacoustic framing
• black-box device evaluation design
• dummy-head recording rig
• Stack: MATLAB R2021a, Python, Neumann KU 100 dummy head, Behringer U-PHORIA UMC204HD USB audio interface, four commercial ANC headphones (Bose QuietComfort 25, Sennheiser HD 4.50BTNC, Sony WH-1000XM3, Sony WH-1000XM4)

Eighty listeners, 360 degraded sentences each, and a mostly negative result

• Speech-chimera synthesis (17-band Hilbert ENV/TFS decomposition with dc-bias depth control)
• repeated-measures counterbalanced design
• CRM closed-set intelligibility task
• rationalised arcsine transform
• IQR-based outlier rejection
• NASA-TLX six-dimension effort survey
• two-way ANOVA
• Stack: MATLAB with Psychophysics Toolbox 3, Ubuntu 20.04 LTS, M-Audio M-Track 2x2 USB interface, CRM corpus (8 talkers), four commercial ANC headphones

Ten instrumental metrics, two corpora, and a jury that cannot agree

• Instrumental audio/speech quality evaluation
• full-reference and no-reference metric application
• modulation-depth function estimation
• cross-corpus replication
• two-way ANOVA with partial eta-squared
• Stack: MATLAB, reference implementations of ten published metrics used as-is, CRM corpus (260 signals, 26 depths), NOIZEUS corpus (990 signals, 33 depths), KU 100 dummy-head recordings of four commercial ANC headphones

A simulated auditory nerve as a virtual listener for ANC headphones

• Computational auditory modelling
• characteristic-frequency and best-modulation-frequency tuning
• rate-MTF
• coefficient of variation of inter-spike interval
• vector strength / synchrony-MTF
• stochastic spike-train averaging
• two-way ANOVA with partial eta-squared
• Stack: MATLAB, custom Python wrapper scripts, AMToolbox 1.0.0 (Zilany 2014 auditory-nerve model), UR EAR toolbox 2020b (SFIE band-enhanced/band-suppressed IC cells), KU 100 dummy-head recordings of four commercial ANC headphones

Building a modulation transfer function for an ANC headphone

• AM sinusoid synthesis
• Hilbert envelope extraction
• correlation-based time alignment
• modulation transfer function estimation
• eight envelope-distortion metrics
• two-way ANOVA with partial eta-squared effect sizes
• Stack: MATLAB R2021a (Signal Processing Toolbox, Audio Toolbox — acousticLoudness/acousticSharpness per ISO 532-1), AMToolbox audspace, Neumann KU 100 dummy head, Behringer U-PHORIA UMC204HD

Designing and teaching a graduate course from scratch

• Independently designed and delivered eleven lectures of a core MSc cognitive science course, covering machine learning and an added game theory module
• Authored and graded two assignments, two quizzes, the end-semester exam and a twenty-topic course project
• Implemented algorithms from scratch live in class and set assignments requiring students to complete working implementations
• Nominated for an institute teaching award; analysed the anonymous course evaluation data against my own teaching statement

Identifying users by how they type — a functional-link neural network for keystroke dynamics

• Keystroke feature extraction (dwell and flight times, digraphs)
• multi-class classification
• FLANN with trigonometric functional expansion
• ROC/AUROC/EER/FNR@zeroFPR evaluation
• fixed-text vs free-text comparison
• Stack: Python, NumPy, scikit-learn, Flask (proof-of-concept login portal), CMU password dataset, free-text keystroke dataset

Keystroke analysis — user authentication and cognitive processes from typing

• Applied simple neural networks to recognise humans from their typing patterns (keystrokes).
• Implemented neural network from scratch in NumPy.
• Built a proof-of-concept web app for multi-user authentication using keystrokes.
• Explored cognitive processes through keystroke data.

Natural language search interface for an academic research database

• Built a natural-language query interface over a database of 500k+ research papers.
• Coordinated responsibilities across team members; suggested the tech stack.
• Personally built the data-ingestion script and the data schema.

PhD, Cognitive Science   CPI 8.68

MSc, Cognitive Science   CPI 8.85

BSc (Hons.), Computer Science   73.7%