$whoami --verbose
Saif Shikalgar
Building in Embedded & Robotics
I build systems end to end ā from microsecond-precision firmware on bare metal to production apps with real users. Recent work spans an automotive intrusion-prevention firewall, a published Android app, and a local-first AI image search engine.
01 Ā· Stack
Skills & Core Technologies
6 Core Engineering Pillars Ā· Full-Stack to Bare-Metal
Embedded Firmware & Hardware
Microsecond-precision bare-metal firmware and RTOS driver architecture.
Automotive & Embedded Security
Hardware intrusion prevention, packet filtering and attack mitigation.
Robotics, Autonomous & Controls
Perception-to-actuation pipelines, closed-loop PID and battery telematics.
Android & Mobile Engineering
Production mobile applications with offline storage and native SDKs.
AI & Edge Intelligence
Deep learning inference, vector embeddings and real-time computer vision.
Full-Stack, Cloud & Web3
Scalable web platforms, microservices, containerization and smart contracts.
02 Ā· Stories & Field Notes
Forensic Engineering Logs & Field LessonsExperience from the trenches
Raw technical breakthroughs, midnight hardware bring-up crisis migrations, microsecond assembly optimizations, and 44 production field stories across 6 specialized engineering domains.
The 3.125Ć Clock Frequency Discrepancy & The Silent UART Baud Divisor Lockup
How an uncorrected HSE_VALUE in STM32 HAL calculated a 562.5 MHz phantom clock, skewing delays and baud rates while hardware PLL ran true
During initial bring-up of the inline CAN Firewall on STM32F446RE, the UART debug console returned corrupted character streams at 115,200 baud, and visual LED heartbeat pulses configured for 500 ms took 1.575 seconds. Yet, CAN bit timings (500 kbps) and hardware timer TIM5 microsecond timestamps were accurate. Forensic investigation revealed HSE_VALUE in stm32f4xx_hal_conf.h was left at the default 25 MHz instead of the physical 8 MHz crystal (a 3.125Ć discrepancy), causing software clock derivation to compute a phantom 562.5 MHz core clock.
The Bidirectional State Collision & Complete Reverse-Channel Starvation
How sharing a single policy timestamp state across dual CAN controllers dropped 98.9% of legitimate reverse-channel frames
In our dual-bus CAN gateway between untrusted and protected buses, bidirectional diagnostic and control CAN IDs (e.g. 0x130) forwarded forward traffic at 94.0%, but dropped 98.9% of reverse frames (only 1.1% delivered). The statistical rate-limiter stored only one timestamp per policy ID, comparing reverse responses against recent forward frames and misclassifying them as DoS bursts.
Why 4D Math Choked Our CPU (And Why Simpler is Better)
Scrapping an Isolation Forest for a 2-Feature Mahalanobis Distance that executes in 56 cycles on Cortex-M4 FPU
When building our CAN bus intrusion firewall Gate 2, the initial 4-feature Isolation Forest burned 2,284 cycles (~12.7 Āµs) per frame. Feature ablation across 988,844 frames proved entropy was useless (moved AUC by <0.002) and timing had benign jitter tails. Scrapping the ensemble tree for a 2D Mahalanobis Distance (Payload L2 Delta + Burst Density) executed in 56 cycles on the Cortex-M4 FPU (832 cycles / 4.62 Āµs total Gate 2), boosting RPM spoofing AUC from 0.30 to 0.988.
The Split State-Update Rule & The Adversarial Baseline Poisoning Flaw
How updating payload baselines with rejected adversarial frames allowed subsequent attack frames to bypass Gate 2
During simulated RPM spoofing attacks, the firewall blocked the first fake frame, but passed all subsequent attack frames as legitimate (blocking dropped to 82.9%). The anomaly detector was updating its reference payload on every evaluated frame. By updating with the rejected payload, the next spoofed frame had zero payload delta from the new baseline. Withholding payload updates on rejected frames while advancing burst history boosted RPM spoofing blocking to 100.0% while keeping legitimate throughput at 97.4%.
The Masquerade Bimodal Dilemma & The Rate-Tier Stand-Down Engine
Resolving the 71.6 percentage-point coin-flip spread when two ECUs transmit the same CAN ID at nominal cadence
When evaluating Gate 1 against masquerade attacks (where a rogue node transmits a legitimate CAN ID offset by half a phase), delivery of legitimate traffic swung randomly between 28.4% and 100.0% (a 71.6 pp spread). Without cryptographic identity in standard CAN frames, deterministic filtering randomly dropped legitimate frames half the time. We solved this with a Rate-Tier Stand-Down engine: on 2x nominal masquerade traffic, the firewall stands down from dropping to preserve 99.9% legitimate command delivery while firing an authenticated Rate Alarm.
The bxCAN Transmit Mailbox Wedging & Auto-Retransmission Trap
How a momentarily disconnected bus caused permanent bridge deadlock due to unacknowledged frame retransmissions
During bench testing, momentarily disconnecting the protected bus DB9 connector permanently locked the firewall. The 3 hardware transmit mailboxes filled with unacknowledged frames because AutoRetransmission was enabled, returning HAL_CAN_STATE_ERROR on all subsequent frames. The bridge never recovered even after reconnecting the cable. We resolved this by implementing a 2ms mailbox abort watchdog that calls HAL_CAN_AbortTxRequest() on the oldest mailbox when all 3 remain full.
03 Ā· Interactive Demos
Hands-On Engineering VisualizersTest the systems live
Interactive 2D visualizers illustrating real-world autonomous emergency braking and embedded automotive intrusion defense.
Autonomous Emergency Braking (AEB) Physics Simulation
Jetson Orin radar perception + STM32 Tandem Master Cylinder (TMC) Linear Actuator (40 Bar Hydraulic).
04 Ā· Selected Work
Selected projects
Flagship builds, then the rest of an active ~13-repository track record.
FiberOpticCalc
Dec 2025 ā presentProfessional PON network design app ā published on Google Play
A native Android application for telecom engineers and field technicians to design and calculate optical power budgets for Passive Optical Networks, combining precision optical-loss math with live geospatial network mapping. Published on the Google Play Store with 5.73K+ installs, 2.09K+ active devices, and a 4.6ā rating.
CAN Firewall
Aug 2026 ā ongoingInline intrusion-prevention system for automotive CAN buses
Abhyuday Racing ā Drive-by-Wire (aBAJA)
Jan 2025 ā present Ā· two seasonsDrive-by-wire & onboard compute stack for the autonomous BAJA (aBAJA) competition
CYO Image Search
Jan ā Jun 2026Local-first semantic image search engine
More builds
Suraksha360
Feb ā Aug 2026Cybersecurity monitoring dashboard + mobile companion app
CottonX
Apr 2026 Ā· Devclash Pune hackathonMulti-agent Web3 orchestration platform
Expert Advisors
Feb ā Jun 2026Algorithmic trading systems for MetaTrader 5
ChainPilot
Apr 2026 Ā· Devclash Pune hackathonAutonomous on-chain agent platform
Floating Translator
Mar 2026 Ā· in progressZero-cost, fully offline on-screen translation tool
Standalone utilities
- ZeroByte
PySide6 desktop GUI for secure, certified disk wiping (drive detection, wipe engine, wipe-completion certificates) targeting NIST/DoD-style sanitization workflows.
- VulnScan
Windows PowerShell system security auditor ā OS/.NET fingerprinting, AMSI provider enumeration, AV/firewall inspection, autorun analysis ā with disk-image test fixtures.
- Email Validator
Node.js/TypeScript CLI wrapping deep-email-validator for regex/MX/SMTP/disposable-address checks.