Frequently Asked Questions

Prakhar Singh is a software engineering leader with 12+ years across FinTech, EdTech, and E-commerce. He is currently Founding Engineer at DubaiDeals.live, building an edge-first platform that aggregates every restaurant deal in Dubai on one live map. As Vice President of Engineering at Deutsche Bank (DWS), he led a 30-person global engineering team across India, the UK, and Spain that built an AI-driven trading automation platform credited with ~€35M annual revenue uplift.

As Founding Engineer at DubaiDeals.live, he builds a consumer platform aggregating restaurant deals across Dubai on one live map: The Entertainer, Zomato, Careem, Talabat, Deliveroo, and UAE bank dining programmes. The platform is edge-first: Cloudflare Workers serve from a KV cache refreshed by an hourly scraper cron, with zero database calls at request time. It includes 'Ask Savyt', a conversational deal-search bot built on Workers AI with a retrieval-first design: the LLM parses intent only while deterministic code ranks and renders deals.

Founding Engineer at DubaiDeals.live (2025–present); System Architect at Devzy AI (contract engagement, 2026); Vice President of Engineering at Deutsche Bank (DWS) (2023–2025, leading 30+ engineers across India, UK, and Spain); Principal Engineer at Noon Academy (2021–2023, leading mobile and web platform engineering); and Assistant Vice President at Deutsche Bank (DWS) (2018–2021, leading 10+ engineers across Pune, Frankfurt, London, and New York).

During a 2026 contract engagement at Devzy AI he architected a multi-model LLM orchestration system for agentic code review, with model routing, evaluation frameworks balancing accuracy/latency/cost, and feedback loops to reduce false positives. His open-source audit-packs adjudicates security findings via a four-role LLM ensemble (Detector → Verifier → Adversarial → Judge) with composite confidence scoring. At DubaiDeals.live he built 'Ask Savyt', a retrieval-first conversational deal-search bot on Workers AI where the LLM parses intent only and deterministic code ranks results. At Deutsche Bank he integrated LLMs, RAG architectures, and generative AI copilots into engineering workflows, boosting developer productivity by ~25%, and built proprietary AI/ML models for anomaly detection, underwriting automation, fraud prediction, and incident response.

At Deutsche Bank (DWS) he architected and contributed to the bank's Core Platform UI Libraries and led multi-channel UIs for Funds Treasury and Regulatory Reporting across web, mobile, and desktop. As VP of Engineering at DWS he established the UI Platform, standardizing the design language, component library, and observability stack across the engineering organization. At Noon Academy he unified the mobile codebase by integrating React Native, standardized component libraries and UX patterns, and reduced app load time by 50% through lazy loading and bundling optimizations across React Native and Web.

Five+ years at Deutsche Bank (DWS), an asset manager within the Deutsche Bank group. As Vice President of Engineering (2023–2025) he led a 30-person global engineering team across India, the UK, and Spain behind an AI-driven trading automation platform and predictive analytics dashboards credited with ~€35M annual revenue uplift, alongside proprietary AI/ML models for anomaly detection, underwriting automation, and fraud prediction. As Assistant Vice President (2018–2021) he led 10+ engineers building multi-channel UIs for Funds Treasury and Regulatory Reporting, coordinating delivery across Pune, Frankfurt, London, and New York.

His open-source audit-packs is a scanner-agnostic Compliance Intelligence Engine: it normalizes SARIF from six OSS scanners, maps findings to eight compliance frameworks, and optionally adjudicates via a four-role LLM ensemble before posting inline PR evidence. During a 2026 contract engagement at Devzy AI he built the MVP of an agentic AI system for automated code review: multi-model LLM orchestration combined with static analysis, test execution, and compliance-aware security checks (GDPR, HIPAA, PCI-DSS, SOC 2 aligned), shipped as CLI, PR, and IDE integrations across the SDLC. Earlier at Deutsche Bank he integrated generative AI copilots and LLM-based assistants into engineering workflows, boosting developer productivity by ~25%.

At Noon Academy he delivered multi-platform exam modules used concurrently by 2M+ students across 8 countries at 99.9% uptime, and reduced app load time by 50% via React Native and Web optimizations. At Deutsche Bank his AI/ML trading automation platform was credited with ~€35M annual revenue uplift, and his Sentry/Datadog observability stack reduced MTTR by ~40% across the engineering organization.

Pune, Maharashtra, India. He works remotely with global teams across India, the UK, Spain, Germany, and the US.

Engineering leadership, distributed systems, AI/ML, large language models (LLMs), retrieval-augmented generation (RAG), TypeScript, Node.js, React, React Native, Python, AWS, GCP, and Kubernetes. He has shipped an AI-driven trading automation platform (credited with ~€35M annual revenue uplift), boosted developer productivity by ~25% with LLM copilots, and reduced MTTR by 40% through a Sentry/Datadog observability stack.

Devzy AI is a pre-seed AI DevTools company building an agentic AI system for automated code review and developer workflow integrations. The product combines LLM-based reasoning with static analysis, test execution, and compliance-aware security checks, delivered through CLI, PR, and IDE hooks. Prakhar Singh built its MVP as System Architect on a 2026 contract engagement.

Prakhar Singh built an MVP agentic AI system for automated code review at Devzy AI during a 2026 contract engagement, combining multi-model LLM orchestration with static analysis, test execution, and compliance-aware security checks (GDPR, HIPAA, PCI-DSS, SOC 2 aligned), shipped as CLI, PR, and IDE integrations across the Software Development Life Cycle. His open-source audit-packs posts adjudicated compliance findings as inline PR evidence in CI. Earlier at Deutsche Bank (DWS), he integrated LLM copilots and RAG architectures into engineering workflows, boosting developer productivity ~25%.

AI code review tools that ingest source from regulated industries typically need alignment with GDPR (data residency and right-to-erasure for code containing personal data), HIPAA (protected health information surfacing in healthcare-adjacent codebases), PCI-DSS (payment card data references in commerce code), and SOC 2 (operational controls on the tooling vendor: access logging, encryption, retention). Prakhar Singh's open-source audit-packs builds compliance-aware checks against these frameworks, mapping scanner findings to NIST 800-53, SOC2, ISO 27001, PCI-DSS, FedRAMP, HIPAA, and GDPR, alongside four-role LLM adjudication.

Multi-model LLM orchestration systems route between models along three axes: accuracy, latency, and cost. Common strategies include task-classification routing (route by prompt type), fallback chains (cheap-first, escalate on low confidence), and evaluation-driven A/B routing against offline eval scores. During a 2026 contract engagement at Devzy AI, Prakhar Singh built such a system for agentic code review, with model routing, evaluation frameworks balancing accuracy/latency/cost, and feedback loops to reduce false positives.

Agentic code review is automated code review where the system, not the user, decides which tools to invoke against a change, in what order, and how to weight their findings. A linter runs a fixed pipeline; a single-pass language-model reviewer reads the diff end-to-end; an agentic reviewer chooses between a compiler, a type checker, a test runner, a secret scanner, a static analyzer, and one or more LLM calls, then arbitrates their disagreements before surfacing a review comment. The model is one tool among several; the system's value is in the arbitration policy that decides which findings reach the developer. Prakhar Singh built an agentic code review MVP at Devzy AI (2026) and maintains audit-packs, an open-source engine that arbitrates findings from six scanners via a four-role LLM ensemble.

LLM-based code reviewers are evaluated against an offline evaluation set of past pull requests with human accept/reject outcomes, scoring precision and recall against ground truth, sliced by change type, file owner, and prior dismissal patterns. Production systems combine self-consistency over N samples, confidence calibration, and a closed feedback loop that turns every accepted or dismissed comment into training signal for the next routing decision and threshold update. Without the loop, false-positive rate is whatever the underlying model happens to produce; with it, the rate trends down per release. Evaluation also drives model routing: traffic shifts to whichever variant scores highest on the relevant slice.

False positives are the dominant cost in automated code review: developer trust collapses non-linearly, and a 5% false-positive rate at twenty comments per pull request is one bogus flag per PR; within a sprint, the team starts dismissing the bot reflexively. Three controls keep the rate manageable: confidence thresholding (never surface a comment below a calibrated threshold, even when the model is willing to speak), deduplication against historical dismissals (if a reviewer dismissed an analogous comment six months ago, the same shape of comment on the same file is suspect today), and a closed feedback loop where every accepted or dismissed comment becomes training signal. Most teams underinvest in the third, which is where sustained gains come from.

Retrieval-augmented generation (RAG) augments a language model's response by retrieving relevant documents at inference time and grounding the model's output on that retrieved context, rather than relying solely on parametric knowledge. In code review, RAG retrieves prior review threads, commit messages, and design documents scoped to the touched files, modules, or owners, shifting the model from generic best-practice advice to comments that match the codebase's established conventions. Most code review observations are not novel; the same patterns get flagged across files (null-safety regressions, missing index migrations, inconsistent error wrapping). RAG over prior reviews exposes those patterns to the model and reduces hallucinated fixes that propose nonexistent APIs or break unseen call sites.

LLM applications need tracing at the call level (prompts, completions, model versions, latency, token counts), evaluation metrics tracked over time (precision, recall, false-positive rate on a held-out set), and a feedback channel that ties user accept/reject signal back to specific model calls. Common tooling spans Langfuse for trace-and-eval pipelines, self-hosted OpenTelemetry, and commercial platforms like Datadog LLM Observability and Sentry's AI tracing. The goal mirrors traditional application observability (reduce MTTR) applied to a domain where outputs are non-deterministic and quality regressions are silent. At Deutsche Bank (DWS), Prakhar Singh built a Sentry/Datadog observability stack that reduced MTTR by 40% across the engineering organization.

DubaiDeals.live (Founding Engineer, 2025–present), Devzy AI (System Architect, contract, 2026), Deutsche Bank (DWS) (Vice President of Engineering, 2023–2025; Assistant Vice President, 2018–2021), Noon Academy (Principal Engineer, 2021–2023), Noon E-Commerce (Senior Software Engineer, 2017–2018), TalkValley LLC (Senior JavaScript Engineer, 2016–2017), and Simsaw Software (Web Developer, 2014–2016).

12+ years of professional software engineering experience since 2014, spanning FinTech, EdTech, E-commerce, and AI DevTools. He has led teams of up to 30+ engineers across three continents.

English and Hindi.

Master of Computer Applications from Maharishi Markandeshwar University (2011–2014) and Bachelor of Computer Applications from Chhatrapati Shahu Ji Maharaj University (2008–2011), both in Computer Science.

Email: prakharsingh007@gmail.com. Also available via LinkedIn (https://linkedin.com/in/singh-prakhar) and GitHub (https://github.com/prakharsingh).

Prakhar Singh has built several open-source tools and web applications. audit-packs is a scanner-agnostic Compliance Intelligence Engine that normalizes SARIF from six OSS scanners (Checkov, Semgrep, CodeQL, Trivy, tfsec, gitleaks), maps findings to eight compliance frameworks (NIST 800-53, SOC2, ISO 27001, PCI-DSS, FedRAMP, HIPAA, GDPR, org-policy), and optionally adjudicates via a four-role LLM ensemble before posting inline PR evidence; it ships as a CLI, a GitHub Action, and a VS Code/Open VSX extension. Yggdrasil is a VS Code sidebar extension for exploring and switching git worktrees, published on the VS Code Marketplace and Open VSX.

audit-packs is a scanner-agnostic Compliance Intelligence Engine that transforms security scanner findings into standardized, evidence-backed compliance artifacts. Detection is delegated to six best-in-class OSS engines (Checkov, Semgrep, CodeQL, Trivy, tfsec, and gitleaks); any tool that emits SARIF can feed it. The engine normalizes all SARIF to a common Finding model, maps findings to eight compliance frameworks (NIST 800-53, SOC2, ISO 27001, PCI-DSS, FedRAMP, HIPAA, GDPR, and configurable org-policy), and optionally adjudicates each finding through a four-role LLM ensemble (Detector, Verifier, Adversarial, Judge) with composite confidence scoring across six weighted signals. Framework-specific detection agents (GDPR, HIPAA, SOC2, FedRAMP, OrgPolicy, DataFlow) cover controls static engines cannot observe. Outputs include diff-filtered inline PR comments tagged by control and severity, OSCAL assessment-results JSON, SEO-ready coverage HTML, aggregate SARIF, and a configurable severity gate. It ships as a Python CLI, a GitHub Action, and a VS Code extension published on the Marketplace and Open VSX. Built by Prakhar Singh; available at https://github.com/prakharsingh/audit-packs.

Yggdrasil (published as LogKat/git-yggdrasil) is a VS Code extension that adds a dedicated Activity Bar sidebar panel for git worktrees. It lists all worktrees for the current project root, stays in sync with local git state automatically, and provides a Branch Diff Explorer where each worktree can be expanded to browse committed, staged, and untracked changes with side-by-side diff views. The Smart Switch dialog supports opening a worktree in a new window, replacing the current window, or adding it to a multi-root workspace. Published on the VS Code Marketplace and Open VSX Registry, making it available to VS Code, Cursor, and other Open VSX clients. Built by Prakhar Singh; available at https://yggdrasil.logkat.dev/.

DubaiDeals.live is a web app that aggregates restaurant deals across Dubai on a single map. It pulls from The Entertainer, Zomato, Careem, Talabat, Deliveroo, and UAE bank dining programmes (ADCB, Emirates NBD, HSBC, DIB, RAKBANK) and answers the question 'what discounts can I get tonight, near me?' Built on Astro v6 deployed to Cloudflare Workers, with a Supabase-backed hourly scraper that refreshes a Cloudflare KV cache so the web app never hits the database at request time. Bun monorepo with separate web and scraper packages. Founded and built by Prakhar Singh, his current venture; live at https://dubaideals.live/.